fix(hicache): discover SGLang context-parallel shards

[ketor]

Summary

• discover SGLang attention-CP and DCP coordinates from the initialized runtime context when dynamic HiCache extra_config omits them
• include the physical CP coordinate in every pool key instead of aliasing all ranks at pcp0/dcp0
• elect one replicated-MLA writer per physical PCP/DCP shard rather than one global TP-rank-zero writer

Why

GLM-5.3-Flash TP8/attention-CP8 produced identical dfkv keys on all eight CP ranks. A restarted target observed some existence hits but no contiguous prefix (prefix=0/1023), issued zero GETs, and recomputed the entire prompt.

Validation

• standalone HiCache key/coordinate suite: 16/16 passed
• xb01-0064, 8x B200, GLM-5.3-Flash, SGLang PCP8 + MTP5 + dfkv, C10 / 64K input / 500 output / 100 prompts
• before: zero dfkv GETs after restart
• after: 522,060 GET hits, 217,627,889,664 bytes read, zero GET/PUT I/O errors
• cold total throughput 24,850 tok/s; restart+dfkv hot 40,763 tok/s (+64%)