Handle unhashable transformer stages in linear precompute prefix detection

[Copilot]

Linear experiment prefix detection could raise TypeError: unhashable type when stage comparison hit transformers that define equality but are not hashable (e.g., LateRetriever-like cases). In that failure mode, _identifyCommon() should skip prefix extraction and proceed with the original pipelines.

• Prefix detection fallback

  • Updated pyterrier/_evaluation/_exec_linear.py so _identifyCommon() catches the unhashable-stage TypeError from _common_prefix(...) and returns (None, pipes).
  • Preserves existing behavior for other errors by re-raising non-unhashable TypeErrors.

• Regression coverage

  • Added tests/test_exec_linear_common_prefix.py with an equality-only (unhashable) transformer to assert graceful fallback instead of failure.

try:
    common_prefix, suffices = _common_prefix(pipe_lists)
except TypeError as e:
    if "unhashable type" in str(e):
        return None, pipes
    raise