[KMCompiler]feat(conj_physical): add hygon runtime and ascend tune config

[tspyc072]

Summary

This PR optimizes conj_physical operator performance on Hygon BW1000 and Ascend 910B, while verifying that the generic implementation already achieves satisfactory performance on Iluvatar, Metax, and Thead platforms.

Changes

Hygon BW1000

• Add platform-specific conj_physical runtime override with fixed BLOCK_SIZE=256
• Add tuned configs in tune_configs.yaml

Ascend 910B

• Add conj_physical tuned configs in tune_configs.yaml
• Generic implementation is used (no runtime override needed)

Iluvatar / Metax / Thead

• Verified that generic implementation performance is on par with torch
• No platform-specific runtime override needed

Benchmark Results

Iluvatar (generic implementation)

Shape	torch.float16	torch.float32	torch.bfloat16	torch.int16	torch.int32	torch.complex64
[256]	1.016	1.000	1.000	1.000	1.000	1.039
[2048, 2048]	1.000	1.021	1.000	1.021	1.021	1.191
[128, 512, 256]	1.000	0.980	1.000	1.000	1.000	1.210
[32, 64]	1.000	1.000	0.980	1.000	1.000	1.533
[512, 1024]	1.000	0.980	1.000	1.000	0.980	1.093
[2, 3, 4]	0.980	1.000	1.000	1.000	0.980	1.028

Thead (generic implementation)

Shape	torch.float16	torch.float32	torch.bfloat16	torch.int16	torch.int32	torch.complex64
[256]	1.000	1.000	1.000	1.000	1.000	1.675
[2048, 2048]	1.000	1.000	1.000	1.000	1.000	1.017
[128, 512, 256]	1.000	1.000	1.000	1.000	1.000	1.001
[32, 64]	1.000	1.000	1.000	1.000	1.000	1.217
[512, 1024]	1.000	1.000	1.000	1.000	1.000	1.091
[2, 3, 4]	1.000	1.000	1.000	1.000	1.000	1.700

Hygon BW1000 (with runtime override)

Shape	torch.float16	torch.float32	torch.bfloat16	torch.int16	torch.int32	torch.complex64
[256]	1.000	1.000	1.000	0.849	1.200	0.984
[2048, 2048]	1.000	1.000	1.000	1.000	0.833	0.931
[128, 512, 256]	1.000	0.843	0.980	1.019	1.200	0.947
[32, 64]	1.020	1.000	1.000	0.981	1.000	0.984
[512, 1024]	0.843	1.000	1.000	1.178	0.981	1.023
[2, 3, 4]	1.000	1.000	0.981	1.000	1.000	0.944

Metax (generic implementation)

Shape	torch.float16	torch.float32	torch.bfloat16	torch.int16	torch.int32	torch.complex64
[256]	1.000	0.974	1.000	1.049	1.044	1.058
[2048, 2048]	1.026	1.026	1.023	0.925	0.976	1.034
[128, 512, 256]	1.000	1.022	1.027	1.024	1.027	0.991
[32, 64]	1.000	0.953	1.054	1.028	1.000	1.115
[512, 1024]	0.974	1.054	0.974	1.000	0.864	0.951
[2, 3, 4]	1.027	1.000	1.000	0.848	1.098	1.038

Tests

• benchmark/test_conj_physical.py passed on all tested platforms
• Pre-commit checks passed

Notes

Platform	Strategy
Iluvatar	Generic implementation
Thead	Generic implementation
Metax	Generic implementation
Hygon	Runtime override + tune config
Ascend	Tune config only (generic implementation)