perf: optimizing collect_node_signatures function

[seroperson]

Closes #966

Summary

This PR introduces set of optimizations to the collect_node_signatures:

• Recursive walk_node approach is replaced with TreeCursor (tree.walk()) and a plain while loop.
• Leaf ("leaf", node_type, text) replaced with ("leaf", node_type, bytes) to strip unnecessary utf-8 decoding.
• Some micro-optimizations, like double node.text call, list allocations, and sticking to C calls.

Benchmark

Benchmarking e2e scoring:

  | fixture     | before (HEAD~1)   | after (this PR)   | speedup |
  |-------------|-------------------|-------------------|---------|
  | py-tiny     |     9.61us        |     8.39us        | 1.15x   |
  | py-small    |    87.65us        |    67.23us        | 1.30x   |
  | py-medium   |     2626us        |     2030us        | 1.29x   |
  | py-large    |    10827us        |     8192us        | 1.32x   |
  | rs-tiny     |    22.35us        |    18.24us        | 1.23x   |
  | rs-small    |   122.10us        |    97.90us        | 1.25x   |
  | rs-medium   |     2193us        |     1700us        | 1.29x   |
  | rs-large    |     9024us        |     6527us        | 1.38x   |
  | sh-tiny     |     7.68us        |     6.32us        | 1.21x   |
  | sh-small    |    76.20us        |    63.66us        | 1.20x   |
  | sh-medium   |     1633us        |     1198us        | 1.36x   |
  | sh-large    |     6595us        |     4775us        | 1.38x   |

python -m benchmarks.bench_collect_node_signatures (run in the branch with actual changes)

benchmarks/bench_collect_node_signatures.py

Test plan

Added tests/validator/test_collect_node_signatures.py which pins the exact function behavior, and also added more e2e tests to the tests/validator/test_token_scoring_integration.py to pin even more the existing scoring behavior (to not break anything in future).

[anderdc]

good stuff