feat(provider): add maxInputTokens window override and maxTokensAsOutputReserve

[nam278]

Summary

• add deepseek-copilot.maxInputTokens — overrides the context window (input + output) advertised to VS Code; 0 keeps each model's configured window (1M for DeepSeek V4)
• add deepseek-copilot.maxTokensAsOutputReserve — when enabled, maxTokens is also advertised as the output reserve, so the advertised input budget becomes window − maxTokens
• cap the advertised output reserve at half the window so advertised output never exceeds input, even when the override is smaller than the model's maximum output
• refresh the model picker immediately when any token-limit setting changes

Why

Copilot Chat computes auto-compact thresholds against the advertised maxInputTokens (background ~0.78, emergency ~0.90). The advertised limits were fixed model constants (input 655,360 / output 393,216), so history kept growing until compaction kicked in near the full 1M window — well past the region where DeepSeek V4 reliably uses its context. At that depth, long conversations degrade: the model starts losing fidelity on older turns and drifting ("lost-at-the-edges" effect), while every request carries hundreds of thousands of tokens of history.

This override lets users set the effective window to ~200K: compaction then kicks in around 156K, keeping the conversation inside the reliable region, with fresh summaries instead of a bloated tail of raw messages. #225 asks for exactly this control.

User impact

• recommended recipe to combat lost-at-the-edges degradation: enable maxTokensAsOutputReserve, set maxTokens: 65000 and maxInputTokens: 265000 → advertised input 200,000, so auto-compact triggers at ~156K and the conversation stays inside the reliable context region
• maxInputTokens: 200000 alone compacts even sooner (~78K trigger) for users who want aggressive cost control
• users who prefer deep context keep 0 and get the previous behavior (advertised input 606,784 / output 393,216, ~473K trigger)
• max_tokens sent to the API is untouched — only the advertised values change

Validation

• npx tsc --noEmit — exit 0
• npm run lint (oxlint) — 0 warnings, 0 errors
• npm run format:check — pass
• hand-verified advertised combos: defaults → 606,784 / 393,216 (trigger ~473K); 200000 → 100,000 / 100,000 (trigger 78K); 265000 + 65K reserve → 200,000 / 65,000 (trigger 156K)

Closes #225