[Nvidia][TritonGPU] Add phases to RemoveLayoutConversions (#763)

* [Nvidia][TritonGPU] Add phases to RemoveLayoutConversions

* [Nvidia][TritonGPU] Gate RLC phases behind a single CMake master switch

* [Nvidia][TritonGPU] Gate RLC phases via compiler.py env switches

* [Sunrise] Inherit pass constructors in RemoveLayoutConversions spec override

* [Nvidia][TritonGPU] Gate RLC enhancement via env switch and compile-time macro

Move the four RemoveLayoutConversions phases behind a runtime master switch
FLAGTREE_RLC_ENHANCE (compiler.py, NVIDIA only, on by default) and wrap all
enhancement code in the __FLAGTREE_RLC_ENHANCE__ compile-time macro (enabled
via CMake for the default/NVIDIA build). Undefining the macro removes the
enhancement and reverts the pass byte-for-byte to the upstream implementation.

Revert the Passes.td pass options and the sunrise spec override back to
upstream; the per-phase switches now live as file-local constexpr flags for
internal debugging only.

* [Nvidia][TritonGPU] Move RLC master switch to knobs; tidy macro guards

Address review feedback on the RemoveLayoutConversions enhancement:

- Drive the runtime master switch through the idiomatic knobs mechanism:
  add nvidia_knobs.rlc_enhance (FLAGTREE_RLC_ENHANCE, on by default) and read
  it in compiler.py, dropping the ad-hoc env-var helper.
- Normalize the __FLAGTREE_RLC_ENHANCE__ / __TLE__ conditional blocks:
  annotate #endif/#else with their controlling macro, restore normal blank-line
  separation between functions, and fix include grouping.

Comment/whitespace/structure only; the macro-on expansion is code-identical to
the previous commit and macro-off still reverts to upstream behavior.

* [Nvidia][TritonGPU] Make resolveConflicts #ifdef branches self-contained

The conflict-resolution loop interleaved the __FLAGTREE_RLC_ENHANCE__ #ifdef
with nested if/for bodies, so the shared closing braces meant different things
in each branch and were hard to follow. Split the loop body into two fully
self-contained #ifdef/#else branches (each with matched braces), duplicating
only the two trivial common lines.

No behavior change: macro-on expansion is code-identical and macro-off still
matches upstream byte-for-byte in this region.

Author: zeroherolin

CMakeLists.txt

@@ -35,6 +35,7 @@ endif()
 if(NOT FLAGTREE_BACKEND)
   add_definitions(-D__NVIDIA__)
   add_definitions(-D__AMD__)
+  add_definitions(-D__FLAGTREE_RLC_ENHANCE__)
 elseif(FLAGTREE_BACKEND STREQUAL "iluvatar")
   add_definitions(-D__ILUVATAR__)
   set(FLAGTREE_TLE OFF)