Add Max Context Length control to instance launch Advanced Options

[alytaphoenix]

Summary

Adds a "Max Context Length" control to the instance-launch Advanced Options panel (alongside Sharding Strategy / Interconnect / Minimum Devices), letting a user cap a model's context window below its native max before launching.

• Model list now carries context_length (already returned by /models, just wasn't in the frontend's local type).
• New numeric input, defaulting to and clamped to the selected model's native context_length; a "reset" link appears once reduced.
• The chosen cap is sent as max_context_length on both the placement-preview request (/instance/previews?...&max_context_length=...) and the /place_instance fallback body used when no precomputed preview is available.
• Resets to "native" whenever a different model is selected (a cap chosen for one model isn't meaningful for another).

Depends on #2240

This is the frontend half of the context-window feature; the backend half (max_context_length on PlaceInstanceParams, get_placement, get_placement_previews, and the placement/memory-validation logic) is in #2240, not yet merged.

Merge-order note: these two PRs are safe to merge in either order. If this one lands first, main's backend simply ignores the extra query param/body field (FastAPI drops unrecognized params silently) — the control renders and clamps, but has no effect on placement until #2240 lands. If #2240 lands first, this PR's control immediately becomes functional.

Test plan

• npx svelte-check — same 15 pre-existing errors / 6 warnings as main (no new errors introduced; confirmed by diffing against a main baseline run)
• npm run build — succeeds
• Manual verification via headless Playwright against a local uv run exo (see caveat below):
  • Selecting a model pre-fills the input with its native context length (e.g. 262,144 for a 256K-context model)
  • Reducing the value triggers a re-fetch of /instance/previews with &max_context_length=<value> in the request, confirmed by intercepting the network request
  • Entering a value above the native max clamps the input back down to it
  • A "reset" link appears once the value differs from native, and clears the cap when clicked
• Caveat: this branch is off main, which doesn't yet have Add configurable context-window with memory-safety validation #2240's backend changes — so this verification covers request formation and input clamping only, not the cap's actual effect on placement or memory validation (that requires Add configurable context-window with memory-safety validation #2240 merged).
• nix fmt / nix flake check not run in this environment; ran prettier directly on the touched files instead.

🤖 Generated with Claude Code

https://claude.ai/code/session_011rjSfwDBTkmySmfU6NgHKF