[Model] Add KERV optimized runtime operators (#1291)
Summary

Bundle the KERV runtime optimization package under
flagscale/models/kerv/ops.
Register 18 torch.ops.flagos_embodied interfaces and integrate the
14-interface KERV runtime profile.
Route KERV training and inference entrypoints to the bundled operators
before any external source tree.
Add correctness tests, packaging checks, operator metadata,
documentation, and an A100 microbenchmark.

Operator status

14 interfaces are integrated into the KERV runtime profile.
3 interfaces remain explicitly experimental and are disabled by
default.
1 compatibility interface is retained for Value Cache Store.
The current A100 auto routes enable the optimized Static Tree
Attention and RoPE-KV Store paths after correctness and performance
gating. Other interfaces retain their integrated reference/native paths.

Validation

KERV unit suite on A100: 47 passed.
Official pre-commit hooks: all passed.
git diff --check: passed.
Wheel build and isolated installed-package import: passed; all 18
operator schemas register successfully.
Full KERV model initialization reaches the optional LIBERO evaluator;
the evaluator itself is kept outside CI because it requires a host
EGL/MuJoCo rendering environment.

Contributors

Zihao Zheng (@zhengzihaoPKU)
Zhihao Mao (@lusunn111, htxmzh@gmail.com)

Co-authored-by: Zhihao Mao 49109638+lusunn111@users.noreply.github.com

Co-authored-by: Zhihao Mao 49109638+lusunn111@users.noreply.github.com