refactor: improve hand divider performance a bit

[Nihisil]

Some Python related optimizations were made, algorithm remains unchanged.

Summary of changes:

• Custom __eq__/__lt__ methods were replaced by @dataclass(frozen=True, order=True)
• Changed _BlockType to (int, Enum) to enable comparison
• Reordered fields to (tile_34, ty) to maintain correct sort order
• from_meld now uses meld.type directly instead of is_chi/is_pon/is_kan
• Added remaining count tracking to avoid sum() at terminal nodes in recursion

All 2.1kk hands in validation set processed without errors.

Benchmark results

Old

Throughput: 28443 hands/sec (based on median)
Avg per hand: 0.035ms (based on median)

New

Throughput: 29315 hands/sec (based on median)
Avg per hand: 0.034ms (based on median)

Overall profile

Metric	Old	New	Change
Total profile time	16.603s	15.445s	-7.0%
Total profile calls	74,744,352	67,800,659	-9.3%

divide_hand function

Metric	Old	New	Change
Own time	0.368s	0.362s	-1.6%
Cumulative time	5.090s	4.097s	-19.5%
Callee calls	1,760,004	1,760,004	0.0%
Callee tottime	0.807s	0.780s	-3.4%
Callee cumtime	4.722s	3.736s	-20.9%

[Apricot-S]

LGTM