fix(controller): install prebuilt CUDA llama.cpp on Windows with staged, digest-verified swaps

[luminary19]

Summary

Adds a Windows prebuilt-CUDA install path for managed llama.cpp and makes
upgrades survive a locked running binary. On win32, installManagedLlamacpp
now queries the latest ggml-org/llama.cpp release, detects the driver's CUDA
version from the nvidia-smi banner, picks the newest compatible
bin-win-cuda-*-x64 build plus its matching cudart runtime, downloads both,
verifies each asset's sha256 digest, extracts with the bundled
System32\tar.exe, and installs by staged rename swap: extract to a staging
dir, rename the old install aside, rename staging into place, then delete the
retired copy. Because the swap never writes over the live executable, an upgrade
while a model is loaded no longer fails on a locked llama-server.exe; a still
-locked retired copy is detected up front with a clear "stop it and retry"
message. Downloads and extraction are bounded by timeouts.

The driver-CUDA banner parser is shared with the wave-1 cuda-version.ts
extractor (one regex, CUDA (?:UMD )?Version) instead of a second copy. Config
help now resolves through resolveLlamaServerHelpBinary, which falls back to the
managed install location when llama_bin is unset and the binary is not on
PATH, so the help panel works right after a managed install.

POSIX is untouched: non-win32 still requires git/cmake and builds from
source exactly as before. Stacked on wave-1 cuda-umd-banner (#185); rebased
onto main.

Validation

bun run typecheck and bun run lint clean. The managed-llamacpp unit suite
(13 tests — asset selection, CUDA-version compatibility picking, digest gating,
help-binary resolution) passes.

Verified live on Windows 11 (RTX 4070): a real managed install downloaded
llama-b9988-bin-win-cuda-13.3-x64.zip + matching cudart, digest-verified,
extracted, staged-swapped, and the installed llama-server.exe --version
reported version: 9988 (f4253ef96). With no llama_bin set,
resolveLlamaServerHelpBinary then resolved that managed binary and
--help exited 0. Total install ~156 s over the network.

The POSIX branch is unchanged, so Linux CI builds llama.cpp from source as
before.

UI changes

None (the existing Install/Upgrade and config-help flows now succeed on
Windows).

Risks / rollout notes

• Win32 install fetches from api.github.com and the release CDN; behind a
  proxy/offline box it fails with a clear error and the source-build /
  LOCAL_STUDIO_LLAMA_BIN / LOCAL_STUDIO_LLAMACPP_UPGRADE_CMD escape hatches
  still apply.
• Picks the newest CUDA build at or below the detected driver version; if
  the driver banner is unreadable it falls back to the lowest-CUDA build.
• The staged swap needs room for one extra copy of the install under the data
  dir during the swap; the retired copy is deleted immediately after.

Fixes #212.