Expose n_ctx override in HookedTransformer.from_pretrained (issue #1006)

[brainsnog]

Summary

Fixes #1006.

get_pretrained_model_config already accepted n_ctx as a parameter,
but it was invisible to users — silently falling through **kwargs with
no type hint, no IDE autocomplete, and no docstring entry on
from_pretrained. Users had to edit source files directly to change
context length.

This PR surfaces it as an explicit, documented parameter.

Changes

transformer_lens/HookedTransformer.py

• Added n_ctx: Optional[int] = None to from_pretrained signature
• Added n_ctx=n_ctx to the explicit call to get_pretrained_model_config
  (previously relied on silent kwargs passthrough)

transformer_lens/loading_from_pretrained.py

• Extended the existing n_ctx override block to emit a logging.warning
  when the requested value exceeds the model's trained default, informing
  users of potential reliability and memory implications

tests/unit/test_loading_from_pretrained_utilities.py

• Added test_n_ctx_override_reduces_context — verifies override works
  when reducing below default
• Added test_n_ctx_override_larger_than_default_warns — verifies warning
  fires with correct message when exceeding default

Both tests use GPT-2 Small and run on CPU — no GPU required.

Usage

# Reduce context for memory efficiency
model = HookedTransformer.from_pretrained("gpt2", n_ctx=256)

# Extend context beyond conservative default (e.g. LLaMA 3.2 defaults to 2048)
model = HookedTransformer.from_pretrained("meta-llama/Llama-3.2-3B", n_ctx=32768)