Roll to convex-149.2.6: enable aarch64 pointer compression; trim + self-publish CI

No V8 change from v149.2.5 — same 14.9.207.26 base and cherry-pick lineage
(dispatch-handle UAF + store-elimination/regexp fixes). This release exists only
to publish an aarch64 build with pointer compression enabled and to formalize the
fork's CI.

CI (.github/workflows/ci.yml):
- Trim the matrix to exactly what convex consumes: release only, simdutf always
  (deno_core enables v8/simdutf unconditionally), pointer compression on the two
  aarch64 targets that run it (linux = prod Graviton, macos = dev). The combined
  ptrcomp+simdutf variant was never built before; every non-simdutf and
  standalone-ptrcomp build was dead weight convex never downloaded.
- Drop the build-asan, windows-arm64, and crates.io publish jobs (unused here).
- Move the get-convex publish guard onto this workflow (was hand-patched per
  release branch) so tagged builds self-publish their release assets.
- Run linux (incl. the arm64 cross-compile) on the org large-runner and windows
  on windows-large-runner, both 22.04/jammy-matched to the toolchain these steps
  assume; install pkg-config + build-essential explicitly (absent on the
  large-runner image; V8's gn gen needs pkg-config). Clang/format conditions key
  on the target triple, not the runner-label string.
- Skip the test step for the aarch64-linux cross target: a pointer-compressed V8
  can't run under qemu-user (cage reservation segfaults emulation). The arm64
  ptrcomp build is exercised natively by the aarch64-apple-darwin job.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: gautamg795