build: declare fla-core (flash-linear-attention) as the Kimi-K3 KDA dependency

[Andrewxu313]

Description

Declare fla-core (flash-linear-attention) as a dependency and document it:

• requirements.txt: add fla-core>=0.5.0 (feeds install_requires, so pip install ., pip install -r requirements.txt, and install_deps.sh all pull it).
• docs/INSTALL.md + docs/deploy-kimi-k3-h200.md: note it as the Kimi-K3 / Kimi-linear KDA dependency.

Motivation

Kimi-K3 / Kimi-linear do a bare top-level import of fla in batchgen/models/moonshotai/kimi_linear/model.py (from fla.modules import FusedRMSNormGated, ShortConvolution; from fla.ops.kda import chunk_kda, fused_recurrent_kda) for Kimi Delta Attention. It was not declared in requirements.txt, install_deps.sh, or the docs — so a clean install can import-error on K3. The >=0.5.0 floor matches the requirement asserted by the K3 KDA test-script (chunk_kda must accept use_beta_sigmoid_in_kernel; the model fail-fasts on older fla). NOTE: pinned as a floor rather than == because it is a feature floor; set an exact == to the validated version if a reproducible lock is preferred.

Type of Change

• model — add/extend model support (models/** + registration seam + model kernels only)
• kernel — add/optimize a compute kernel (batchgen_kernels/** + in-tree kernel dirs)
• core — change scheduling/serving/runtime scaffolding (the only type that may)
• fix — narrow bug fix (+ a regression test)
• infra — build / CI / packaging / scripts / Docker
• docs — documentation only

File changes

File	Δ	Note
requirements.txt	mod	add fla-core>=0.5.0 (Kimi-K3 KDA dependency)
docs/INSTALL.md	mod	prerequisite note for the K3 KDA fla-core dependency
docs/deploy-kimi-k3-h200.md	mod	§3 install note for fla-core

Checklist

• I have read the CONTRIBUTING guide and the PR Merge Policy Contract.
• I have updated the tests (if applicable).
• I have updated the documentation (if applicable).