feat(win): fix silent CPU fallback, launcher suite, DirectStorage expert loads

[khalilswdp]

**Stacked on #667
#667 ** (model-architecture seam) — its commit shows here until
it merges. This PR's own work is three commits, reviewable independently:

1. fix(win): static cudart + launcher/cap-warning fixes
2. feat(win): PowerShell launcher suite
3. feat(win): DirectStorage expert-load transport (opt-in)

Summary

On Windows every run silently fell back to CPU: coli_cuda.dll linked dynamic
cudart, but CUDA 13.x keeps cudart64_*.dll in bin\x64 (not on PATH), so
LoadLibrary failed quietly. Fix: -cudart static (the DLL then imports only
KERNEL32). The rest makes the fix visible and usable:

• coli launcher: banner prints the real backend instead of a hardcoded
  "streaming CPU"; numeric env vars parse tolerantly (CUDA_EXPERT_GB=auto no
  longer crashes); --ram accepts fractional budgets.
• cap=1 warning: quotes the actual RAM shortfall and what --ram would
  buy a second cache slot, instead of unconditionally blaming PIN_GB.
• windows/*.ps1 suite: Build (toolchain detect + doctor), Run (wizard,
  attaches to a running serve instead of starting a second engine), Chat, Web,
  Warmup. Saved with a UTF-8 BOM — PowerShell 5.1 reads BOM-less files as
  ANSI, which breaks the multilingual warmup prompts.
• warmup.ps1: -Backend auto|gpu|cpu (the .coli_usage pin is
  backend-flavoured), -PromptFile, +40 multilingual/code prompts.
• DirectStorage expert loads (opt-in COLI_DSTORAGE=1, default off): the
  miss path's coalesced ~19 MB slab pread can be served by a DirectStorage
  host-destination DMA instead. Pure transport swap — bytes land in the slab
  the expert cache already owns; cache/eviction/layout untouched; primary
  replica only; any failure (no SDK at build, no dstorage.dll, no adapter,
  request error) falls back to pread. The depot PR reuses this transport for
  disk→VRAM. No performance claims yet — correctness-verified only.

Validation

• make -C c check full pass at this tip; zero new warnings
• make CUDA_DLL=1 colibri.exe clean; DLL builds with and without
  DSTORAGE_HOME
• DirectStorage A/B on a real int4 container: COLI_DSTORAGE=0 vs 1
  byte-identical output; served loads reported at exit
• Token-exact oracle with the flag off (default): 32/32 TF, 20/20 greedy
• All six .ps1 scripts parse under Windows PowerShell 5.1 and PS 7
• Static-cudart DLL: dumpbin shows KERNEL32 imports only; GPU engages on a
  box where the dynamic link failed silently

Compatibility

• Default CPU build unchanged and dependency-free — DirectStorage is
  opt-in at build time (DSTORAGE_HOME), optional in the loader;
  Linux/macOS unchanged
• No model files or binaries included
• Defaults changed: none in the engine.

[JustVugg]

Holding this behind #667 (its arch-seam base). Please rebase #667 first onto latest dev, then this one on top — CI green on both, just needs the merges resolved.

[JustVugg]

dev has moved a fair amount (v1.3.0: Kimi K3 and Inkling engines, the ragged-attention grouped-scale fix, CUDA_RELEASE_HOST default, the Inkling expert-cache fix). This PR now conflicts — could you rebase onto latest dev? Nothing is wrong with the change itself; CI was green where it ran. Ping me once it's clean and I'll review/merge.