fix/feat!: LLMs context management

[mateuszlampert]

Description

This PR fixes few bugs related to the LLMs, caused by mixing two approaches - functional (as we pass whole messages history each time) and stateful (as we keep pos_ in the runner, representing at which position the KV cache is), which resulted in 3 bugs:

• broken KV cache for reasoning models - in the runner, we counted tokens generated for the reasoning and included these in KV cache (pos_ += num_generated_tokens), but in next turns, jinja template removed these reasoning tokens from the messages history - as a result, KV-cache was incoherent
• duplicated tokens in KV cache - we were passing whole messages history to the runner (functional approach), but we were also appending all tokens (prompt and generated) to the KV cache (which position is represented by pos_) - as a result tokens were "duplicated" in the KV cache and we were running out of available tokens very fast (exceeding context_window_length)
• stateful TS functional API - even though our generate() method is called functional, it kept internal state in the runner (e. g. pos_)

These bugs were fixed by resetting the runner before each generation, which makes it truly functional - old messages are prefilled and the KV cache can be still used during generation phase.

Additionally, this PR adds ContextStrategy to ChatConfig interface, so now it's possible to define (or use one of already implemented) strategy for managing context (e. g. naive, message count based, sliding window) - it gives us more flexibility and user can decide what's best for their use case. From now on, SlidingWindowContextStrategy is also configured as the default one.

Introduces a breaking change?

• Yes
• No

These changes will not break anything until max number of messages is not modified (I removed contextWindowLength from ChatConfig and replaced it with contextStrategy)

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

Run example llm app, open executorch logs (adb logcat | grep -i "executorch" for example) and see if numbers of tokens are properly aligned and if pos_ is correct.

To test different context management strategies, change contextStrategy in llm app and modify model configuration.

Screenshots

Related issues

#776

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

Position in KV cache, number of prompt tokens and number of generated tokens for both non-reasoning and reasoning models BEFORE changes.

LLAMA 3.2 1B SPINQUANT (without reasoning)

pos_	Prompt tokens	Generated tokens
0	335	269
604=269+335	872	372
1848=604+872+372	1513	CRASH

QWEN 3.0 0.6B QUANTIZED (with reasoning)

pos_	Prompt tokens	Generated tokens
0	309	457
766=309+457	617 (<766!)	192
1575=766+617+192	925 (<1575!)	CRASH

[mateuszlampert]

I will update docs after the code gets approved

[mateuszlampert]

I changed only docs for useLLM and LLMModule - the rest of changes is related to autogenerated api reference changes