Pick the right CUDA wheel per GPU, surface integer versions H3 + clip window select

[imprsnst]

Pick the right CUDA wheel per GPU, and stop warning on cards that work

Closes #33

Two field reports after v1.2.64, plus the clip-length controls a user asked for.

What changed

• arch_list_covers replaces exact matching with CUDA's within-major rule. Directional, so an
  sm_80 A100 against an sm_86-only wheel still warns, and sm_90 still does not cover sm_120.
• The sm_ parse rule is written down: minor is the last digit, major is everything before.
  sm_120 is (12, 0), not (1, 20). Reading it the other way puts the bug on exactly the Blackwell
  parts this is about.
• The Windows probe was rewritten. set /a coercion instead of a piped findstr guard, so a
  driver answering Unknown Error yields 0 rather than string-comparing above 10 and winning.
• --torch-index now replaces torch. Project install first, then an exclusive --index-url
  pass for torch and torchvision last, so nothing can re-resolve over the named index. torchvision
  goes with it: a +cu130 torch beside a +cpu torchvision imports fine and dies at the first
  vision op.
• Existing bad installs get repaired. On the reuse path the installed torch is asked whether it
  covers the card. Gated: only a recognisable +cpu/+cuXXX build is replaced automatically, so a
  ROCm wheel or a nightly is left alone with a message instead.
• --print-torch-index reports what the driver said and which index would be used, and installs
  nothing. One line to paste into a bug report.
• Pre-R580 Blackwell gets cu128, loudly. cu128 is frozen at torch 2.11 but still serves and was
  the first index with sm_120, so that machine has one workable choice. Erroring would be worse
  than what v1.2.64 did.
• .inline-install.json records what the installer intended versus what landed.
• Non-NVIDIA installs now say why no GPU was found and point AMD/Intel users at --torch-index.

Clip controls

• Clip length shows what it resolves to: Trains on 107 frames (4.46s) per clip. The 17n+5 grid
  snaps down, so asking for 5s trains on 4.458s, and that was silent until now.
• New Clip window select, start or end, for footage whose action is at the finish.
• Fixed a wrong claim in that help text: it said clips cost "far more memory", but measured, a clip
  peaks within 0.5GB of a still. Length costs time, not memory.

A windows-latest CI job now runs webui.bat on a real cmd interpreter with a stubbed
nvidia-smi, asserting all five decisions. There was no test workflow at all before, which is how
these shipped. It immediately earned its keep: it caught two bugs in my own rewrite, a missing
call that let a .bat shim hijack the script, and a delimiter that split 12.0, 610.88 so the
driver version was never read.