Memory profiling design (#362)

* Memory profiling design

* Fix CI: add frame pointers to RUSTFLAGS, support aarch64 in unwind bench

The RUSTFLAGS env var in the CI composite action overrides
.cargo/config.toml, which includes -C force-frame-pointers=yes.
Add it to the CI RUSTFLAGS explicitly.

Also replace the crate-level #![cfg(x86_64)] with per-item cfg gates
and add aarch64 inline asm so the benchmark runs on ARM CI too.

* Address CR feedback on memory profiling design

Key changes:
- Replace DashMap with crossbeam-skiplist::SkipMap for liveset (lock-free,
  eliminates deadlock risk inside allocator hook)
- Fix realloc bug: on_dealloc now called after realloc succeeds, not before
- Use RAII ReentrancyGuard struct that clears on drop (panic-safe)
- Split goals: "install-once" and "instruments all threads" are separate
- Add non-goals: memory contents inspection, cryptographic randomness
- Add evolution path for mmap tracking
- Add viewer/analysis as explicit goal
- Contextualize overhead with typical allocator latencies (20-80ns)
- Fix CPU calculation: 0.06-0.1% of one core, not "well under 1ms/sec"
- Add shuttle concurrency testing to test strategy
- MUSL: conditionally compile out if untested rather than shipping untested
- MAX_FRAME_SIZE: explain safety tradeoff, plan to raise with validation
- Clean up conversational language throughout

Author: rcoh

.github/actions/rust-build/action.yml

@@ -27,7 +27,7 @@ runs:
         - name: Build and test
           shell: bash
           env:
-              RUSTFLAGS: "--cfg tokio_unstable"
+              RUSTFLAGS: "--cfg tokio_unstable -C force-frame-pointers=yes"
           run: |
               if [ "${{ inputs.toolchain }}" = nightly ]; then
                   rm -fv Cargo.lock