Comgr hotswap: replace hardcoded AMDGPU VGPR/SGPR granule constants with MC-derived values

[harsh-nod]

Background

amd/comgr/src/comgr-hotswap-b0a0.cpp currently carries two hardcoded per-ISA constants for the GFX1250 B0-to-A0 rewrite policy:

static constexpr unsigned Gfx1250VgprGranuleSize = 8;
static constexpr unsigned Gfx1250SgprGranuleSize = 16;

These are hardware facts per AMDGPU generation (gfx10+ wave32: VGPR granule 8, SGPR granule 16) and are already encoded in LLVM — llvm::AMDGPU::IsaInfo::getVGPRAllocGranule(MCSubtargetInfo) / getSGPRAllocGranule(MCSubtargetInfo) — but that header lives inside the AMDGPU target:

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

Comgr also builds standalone against an installed LLVM, and target-internal headers aren't reliably on that include path. That's the same constraint that kept us out of AMDGPUTargetStreamer::getElfMach() during #2201 review.

Proposal

Pick one of:

1. Promote AMDGPU::IsaInfo granule getters to a public LLVM header. Simplest API-wise. Adds a small public surface (llvm/include/llvm/TargetParser/AMDGPUIsaInfo.h or similar) that wraps the existing target-internal implementation. Upstream LLVM change.
2. Add thin helpers to llvm/Support/AMDHSAKernelDescriptor.h. Keeps the new surface AMDHSA-flavored and on a header Comgr already uses.
3. Have comgr query MCSubtargetInfo::getFeatureBits() + a tiny local family table. Avoids touching upstream, but duplicates logic.

Option 1 or 2 are preferred — option 3 recreates the same knowledge LLVM already has.

Once landed

Drop Gfx1250VgprGranuleSize / Gfx1250SgprGranuleSize (and any future per-ISA equivalents) from comgr-hotswap-b0a0.cpp and populate RewriteConfig::VgprGranuleSize / RewriteConfig::SgprGranuleSize via the public helper, driven by the parsed target ISA.

Context

Called out during PR #2203 review: #2203 (comment). Tracking this here so the constants don't accumulate as more source/target policies land. cc @chinmaydd @lamb-j