[Feature] Enable MambaRadixCache (prefix caching) for KDA (KimiLinearForCausalLM)

[yyq0210]

Summary

KDA (Kimi Delta Attention) backend currently has
support_mamba_cache=False, which means KimiLinearForCausalLM
models cannot benefit from SSM state prefix caching via
MambaRadixCache. This PR enables it by:

1. Setting support_mamba_cache=True in server_args for
   KimiLinearForCausalLM
2. Adding the missing _track_mamba_state_decode,
   _track_mamba_state_extend, and init_forward_metadata
   implementations in KDAAttnBackend
3. Returning the intermediate h tensor from chunk_kda_fwd
   (already computed but previously discarded)

Motivation

With MambaRadixCache enabled, requests sharing a common prefix can
reuse cached SSM states instead of recomputing from scratch. In
testing with Kimi-Linear-48B-A3B-Instruct (TP=2,
--mamba-scheduler-strategy extra_buffer):

• Cold request: 401 prompt tokens, 0 cached
• Subsequent requests with shared prefix: 384 tokens cached, only
  16-17 new tokens need computation

This provides significant TTFT reduction for workloads with shared
system prompts or repeated prefixes.

Why it wasn't enabled before

KDA's forward_decode and forward_extend were missing the state
tracking calls that GDN already has. Additionally,
chunk_kda_fwd discarded the h tensor (intermediate SSM states
at chunk boundaries) via del h, making extend-time tracking
impossible.

The implementation is straightforward since:

• The base class MambaAttnBackendBase already provides
  _init_track_conv_indices, _init_track_ssm_indices,
  _track_mamba_state_decode, and _track_mamba_state_extend
• chunk_gated_delta_rule_fwd_h (shared with GDN) already
  computes h — KDA just needs to stop deleting it
• No Triton kernel changes required

Key difference from GDN

KDA stores conv states with transposed layout: (num_slots, conv_width, qkv_dim) vs GDN's (num_slots, qkv_dim, conv_width).
The implementation accounts for this via:

• Swapping conv_states_shape last two dims so
  _init_track_conv_indices sees correct conv_state_len
• Using .permute(1, 2, 0) instead of .transpose(0, 1) when
  writing tracked conv states

[yyq0210]

fix：#26580

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.