Add streaming autocomplete mode

[FuJacob]

Summary

• Rebases streaming autocomplete onto current main, after KV-cache reuse and AXObserver focus tracking landed.
• Resolves conflicts by preserving per-app enable/disable settings alongside the streaming autocomplete toggle.
• Updates the streaming llama path to await async KV-cache reset calls from main so stream failures do not race the next generation.

Validation

xcodebuild test -scheme tabby -project tabby.xcodeproj -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO CODE_SIGNING_REQUIRED=NO
# ** TEST SUCCEEDED **
# Executed 69 tests, with 0 failures (0 unexpected)

Note: the test process still printed the existing llama.cpp/ggml Metal teardown GGML_ASSERT([rsets->data count] == 0) after XCTest reported success; xcodebuild exited 0.

Linked issues

Refs #11
Refs #13

Risk / rollout notes

• Streaming is behind the menu-bar/settings toggle and defaults off.
• Partial output is only surfaced at stable chunk boundaries through SuggestionStreamChunker so token fragments are not shown directly.
• Runtime reset semantics now follow the async reset contract from main for both non-streaming and streaming llama paths.