refactor(sft): unify KTConfig field names with kt_ prefix, add share_cache_pool, remove dead code

- KTConfig fields all use kt_ prefix matching dict keys — eliminates
  _OLD_TO_NEW mapping and prefix-stripping in wrapper.py
- Add kt_share_cache_pool field, auto-enabled when gradient_checkpointing
  is on (via training_args.py), flows through to C++ cache allocation
- Remove dead checkpoint detection code: in_ckpt_recompute,
  in_ckpt_first_forward vars (assigned but never read), fallback
  _is_in_checkpoint_first_forward() function, unused inspect import
- Remove redundant env var fallbacks in wrapper.py for share_backward_bb
  and share_cache_pool (KTConfig.__post_init__ already handles env vars)
- Simplify layer.py checkpoint logic to single _checkpoint_hook_mode() check

Verified: Qwen3-235B 3-step training on sap4, loss matches baseline
(1.2886 / 1.9824 / 1.377 vs 1.2886 / 1.9766 / 1.3809)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kt-kernel/python/sft/dist_utils.py

@@ -9,7 +9,6 @@ This is a leaf module — no imports from other sft/ submodules.
 
 from __future__ import annotations
 
-import inspect
 from contextlib import nullcontext
 from typing import Any
 
@@ -125,18 +124,6 @@ def _dist_scatter_varlen_from_rank0(
     return local_out
 
 
-def _is_in_checkpoint_first_forward() -> bool:
-    """Best-effort detection for non-reentrant checkpoint first forward."""
-    try:
-        for frame_info in inspect.stack(context=0):
-            fn = frame_info.function
-            file = frame_info.filename or ""
-            if fn == "custom_gradient_checkpointing_func" and file.endswith("checkpointing.py"):
-                return True
-    except Exception:
-        return False
-    return False
-
 
 def _checkpoint_hook_mode() -> str:
     """Infer checkpoint phase from current saved_tensors_hooks top.