Move inline classes into a separate file.
Make serialize() into a basic function, support numeric conversions with base option
Refactor numeric conversions and add some more helper functions for the
build flags.

- gpio module now tracks the known providers (right now, hardware and mcp expander)
- refactored relay struct to use 'Provider' implementing setup,notify,change,boot instead of just BasePin actions
- refactored button module to use gpio provider instead of referencing types itself
- removed dual & stm code from buttons, migrate both to relay module
- added status notify and change callbacks for relayStatus (i.e. 'notify' when relay status was called, but not changed. and 'changed' when it did)  
- relays runtime configuration keys
- relay command now shows configured relays and current & target statuses
- refactor the code using relayStatus(0, blah) under LIGHT_PROVIDER check to use lightState instead
- remove rfbridge code form relay module. implement through a basic state listener in the rfbridge module, depend on RELAY_SUPPORT
- allow to bind rf codes to real relays
- drop tuya-specific lights provider, remove tuya code from relays and lights modules
- integrate tuya via relay listeners and providers, use lights custom provider
- implement channel transitions for tuya. disabled by default, and transition time and step are overridden to 2000 + 100. needs to be set to some value below the total time (i.e. `total transition time / step time == number of steps`, we need to figure out a correct time that serial comms could handle)
- lights custom provider (global, not per-pin) and state listeners
- remove lights code from relay module. implement through providers & listeners in the lights module, depend on RELAY_SUPPORT
- lights per-channel relay provider (unused atm), depend on RELAY_SUPPORT
- refactored channel transition - calculate step only once, make sure time + step values are sane, generate quick transitions with very small delay (10ms hardcoded) for transitions during OFF state i.e. we no longer waste 500ms (or whatever transition time is set to) on boot doing nothing
- transition time + step parameter for the lightUpdate
- report mask parameter for the lightUpdate
- minor fixes across the board

resolve #2222

resolve #2310

- revert unintended change to include 2 bytes as part of empty value
  fixes compatibility with old versions including Embedis and having empty values
- overwrite now empty space when moving
- re-shuffle internal logic to be more strict with cursor oob read / writes
- update comments

* wip based on early draft. todo benchmarking

* fixup eraser, assume keys are unique

* fix cursor copy, test removal at random

* small benchmark via permutations. todo lambdas and novirtual

* fix empty condition / reset

* overwrite optimizations, fix move offsets overflows

* ...erase using 0xff instead of 0

* test the theory with code, different length kv were bugged

* try to check for out-of-bounds writes / reads

* style

* trying to fix mover again

* clarify length, defend against reading len on edge

* fix uncommited rewind change

* prove space assumptions

* more concise traces, fix move condition (agrh!!!)

* slightly more internal knowledge (estimates API?)

* make sure cursor is only valid within the range

* ensure 0 does not blow things

* go back up

* cursor comments

* comments

* rewrite writes through cursor

* in del too

* estimate kv storage requirements, return available size

* move raw erase / move into a method, allow ::set to avoid scanning storage twice

* refactor naming, use in code

* amend storage slicing test

* fix crash handler offsets, cleanup configuration

* start -> begin

* eeprom readiness

* dependencies

* unused

* SPI_FLASH constants for older Core

* vtables -> templates

* less include dependencies

* gcov help, move estimate outside of the class

* writer position can never match, use begin + offset

* tweak save_crash to trigger only once in a serious crash

* doh, header function should be inline

* foreach api, tweak structs for public api

* use test helper class

* when not using foreach, move cursor reset closer to the loop using read_kv

* coverage comments, fix typo in tests decltype

* ensure set() does not break with offset

* make codacy happy again

Change the underlying command line handling:
- switch to a custom parser, inspired by redis / sds
- update terminalRegisterCommand signature, pass only bare minimum
- clean-up `help` & `commands`. update settings `set`, `get` and `del`
- allow our custom test suite to run command-line tests
- clean-up Stream IO to allow us to print large things into debug stream (for example, `eeprom.dump`)
- send parsing errors to the debug log

As a proof of concept, introduce `TERMINAL_MQTT_SUPPORT` and `TERMINAL_WEB_API_SUPPORT`
- MQTT subscribes to the `<root>/cmd/set` and sends response to the `<root>/cmd`. We can't output too much, as we don't have any large-send API.
- Web API listens to the `/api/cmd?apikey=...&line=...` (or PUT, params inside the body). This one is intended as a possible replacement of the `API_SUPPORT`. Internals introduce a 'task' around the AsyncWebServerRequest object that will simulate what WiFiClient does and push data into it continuously, switching between CONT and SYS.

Both are experimental. We only accept a single command and not every command is updated to use Print `ctx.output` object. We are also somewhat limited by the Print / Stream overall, perhaps I am overestimating the usefulness of Arduino compatibility to such an extent :)
Web API handler can also sometimes show only part of the result, whenever the command tries to yield() by itself waiting for something. Perhaps we would need to create a custom request handler for that specific use-case.

- add experimental `relayDummy` to configure dummy relays at runtime
- add tuya-generic-dimmer #1729
- cleanup broker interface to allow Tuya module to properly receive events, modify broker methods to allow different function signatures
- add basic tests for Tuya frame and data protocol