BeamSearchStrategy: silent worker starvation when num_expanding_parents=0

🐛 Describe the bug

num_expanding_parents accepts 0 and negative values without validation.
A value of 0 causes select_candidates() to return an empty candidate list,
spawning zero workers per round. All max_rounds rounds complete silently
with no optimization progress and no error or warning surfaced.

location: triton_kernel_agent/opt_worker_component/searching/strategy/beam_search.py

reproduction:

from triton_kernel_agent.opt_worker_component.searching import BeamSearchStrategy
s = BeamSearchStrategy(num_expanding_parents=0)
s.initialize(program)
candidates = s.select_candidates(round_num=1)
# candidates is empty and zero workers spawn, rounds consumed silently

two related problems I found:

• BeamSearchStrategy candidate dicts are missing the inspirations key
  required by the SearchStrategy Protocol. GreedyStrategy already
  provides this key; the inconsistency means future Mutator implementations
  reading inspirations from the candidate dict would silently get nothing
  under beam search.
• ptx_hash_from_cache silently skips unreadable PTX files with no log
  output, returning an incomplete fingerprint with no indication to the caller.

Have a fix ready for all three with 19 new tests in PR #146. All 72 existing tests
pass

Platform and Version

main branch, tested on Google Colab (NVIDIA T4 GPU), Python 3.12

Start with triton_kernel_agent/opt_worker_component/searching/strategy/beam_search.py, especially BeamSearchStrategy.select_candidates(), then compare its candidate dictionaries with GreedyStrategy and the SearchStrategy Protocol. Locate ptx_hash_from_cache and review PR #146 and its 19 new tests; done means invalid parent counts are handled, inspirations is present, and unreadable PTX files are reported without silent incomplete results.