Implement triton kernels for inference

[yuanheng-zhao]

Tracking for implementation of triton kernels compatible with relevant submodules and KVCache for inference.

• Context-stage Attention [kernel] Add triton kernel for context attention without padding #5192
• Decoding-stage Attention [kernel] Add flash decoding triton kernel for blocked kv cache #5249
• Pos Embedding
  • [Inference] Kernel: no pad rotary embedding #5252
• KVCache Copy
  • context: fused in context attention
  • decoding: [kernel] Add KV cache copy kernel during decoding  #5261
• Layer norm [kernel] Add RMSLayerNorm triton kernel #5262

[CjhHa1]

First stage kernel implementation is done