Include extra specialization in specialization_key

[hinriksnaer]

Addresses #1209

specialization_key() now returns the full specialization key, including constraints discovered during compilation. There might be a more elegant solution to this but would likely involve more risk and potential for introducing bugs. Hoping that this fully addresses #1209.

Changes

• Extracted _base_specialization_key() from specialization_key() for internal use
• specialization_key() appends compilation-discovered extras (e.g. from hl.specialize()) when available
• Updated bind(), local_cache.py, and aot_cache.py to use _base_specialization_key() internally
• Added regression test in test_specialize.py

Note

Before the first compilation, extras are not yet known, so the key may still be incomplete. I am not sure if this would cause any existing autotuning cache results to miss.

[fulvius31]

LGTM but base_search.py should use _base_specialization_key() so that FROM_BEST_AVAILABLE matches across different hl.specialize values and seeds the autotuner with prior good configs instead of falling back to defaults.

It's not the scope of this PR, I'll follow up with another PR when this get merged.

I am not sure if this would cause any existing autotuning cache results to miss.

Yes but I don't think is relevant.