[GPU] some fixes and optimizations to CM PA and XAttention kernels

[ceciliapeng2011]

Improve KVCache quantization, XAttention flexibility, and sparse attention performance.

Details:

• Use float as internal precision for KVCache quantization in kvcache_update CM kernel to fix accuracy issues in QWen3-32B int8 model.
• Remove restriction in PA 2nd token CM kernel that limited heads_num / kv_heads_num <= 8, resolving MiniCPM4 failure.
• Fix phi-3-mini-128k-instruct issue caused by head_size=96 not divisible by 64 in xattention_gemm_qk kernel.
• Support k/v head_size 96 in kvcache_update kernel
• Root cause and fix issue that even with xattn_threshold 100, for both i8/fp16 kvcache, first token of MiniCPM4 still wrong.

Tickets:

• CVS-178816
• CVS-178638