[AMDGPU][COMGR] Install gfx125x entry trampolines with hotswap

[harsh-amd]

ISSUE ID: SWDEV-499597

Summary

• Add gfx125x kernel-entry trampoline rewriting to COMGR hotswap behind AMD_COMGR_HOTSWAP_ENTRY_TRAMPOLINES.
• Keep the entry-trampoline pass separated from the B0-to-A0 instruction patch dispatcher in comgr-hotswap-entry-trampoline.cpp.
• Use llvm::object::ELFFile<> for grown-image ELF symbol traversal when adjusting symbol values, matching the COMGR agent conventions.
• Keep COMGR as the rewrite engine/API only: this PR no longer builds or installs libamd_comgr_hotswap_tool.so.
• Use explicit hotswap-local gfx1250 stepping features so B0-to-A0 instruction patches only run for explicit B0-to-A0 requests, while entry trampolines remain opt-in.
• Update COMGR hotswap docs to point runtime users at rocm-systems libhsa-hotswap.so.

Paired PRs

• Runtime HSA tool: [rocr] Revert loader-side gfx125x entry trampolines rocm-systems#7577
• TheRock integration: feat(hotswap): integrate entry trampoline subprojects TheRock#6018

Testing

• git diff --check
• ASAN_OPTIONS=use_sigaltstack=0 cmake --build build-comgr-asan --target amd_comgr HotswapElfTests HotswapMCTests hotswap-rewrite --parallel 32
• ASAN_OPTIONS=use_sigaltstack=0 build-comgr-asan/test-unit/HotswapElfTests
• ASAN_OPTIONS=use_sigaltstack=0 build-comgr-asan/test-unit/HotswapMCTests --gtest_filter="BuildKernelEntryTrampoline.*"
• ASAN_OPTIONS=use_sigaltstack=0 build-llvm/bin/llvm-lit -sv build-comgr-asan/test-lit/hotswap-elf-growth.s build-comgr-asan/test-lit/hotswap-kernel-entry-trampoline.s build-comgr-asan/test-lit/hotswap-kernel-entry-trampoline-multi.s build-comgr-asan/test-lit/hotswap-barrier-isfirst.s
• ASAN_OPTIONS=use_sigaltstack=0 cmake --build build-comgr-asan --target check-comgr --parallel 32

ASAN check-comgr passed on mi350-4 at e199f875ac611dfa82e4ce7ffb1f4c89e008fa44: lit reported 70 passed / 11 unsupported, CTest reported 31/31 passed, and HotswapElfTests reported 6/6 passed.