refactor: optimize performance for hand calculations

[Nihisil]

I think it is the latest round of optimizations. All low-hanging fruit have been picked (only removal of is_agari remained, but it would be done in separate PR).

We can optimize it further for sure, but it would require more invasive changes to the architecture, and the goal of this lib is not to be the fastest mahjong lib out there, but rather a good balance of speed, correctness, and maintainability. For super fast libs there are Rust/C versions available.

For example, we can get a lot of performance gains by getting rid of Yaku.is_condition_met() and just pre-calculate all honors, suits and etc per hand and use these ints to calculate all yaku, but it is not the change that I want to do in this lib.

Summary of optimization patterns used:

• Algorithmic improvements: O(n^2) → O(n) via dictionary counting and bitmask approach
• Data structure upgrades: list → frozenset/set for O(1) membership, bitmask for intersection
• Pass reduction: Multiple iterations collapsed into single-pass logic
• Some hand yaku pre-checks to avoid not necessary calculations, mainly based on https://riichi.wiki/Yaku_compatibility

And all 2.1kk validation hands passed successfully.

Benchmark Results

Benchmark was run with n=5, limit=200000.

Old

• Throughput: 28313 hands/sec (based on median)
• Avg per hand: 0.035ms (based on median)

New

• Throughput: 30985 hands/sec (based on median)
• Avg per hand: 0.032ms (based on median)

Overall profile

Metric	Old	New	Change
Total profile time	15.691s	14.021s	-10.6%
Total profile calls	67,801,516	56,573,351	-16.6%

estimate_hand_value function

Metric	Old	New	Change
Own time	2.964s	2.462s	-16.9%
Cumulative time	15.645s	13.964s	-10.7%
Callee calls	12,852,409	10,221,010	-20.5%
Callee tottime	5.174s	4.544s	-12.2%
Callee cumtime	12.680s	11.502s	-9.3%

[Apricot-S]

LGTM