perf(translate): fix slow translation startup — CJK estimation, output budget, provider token config

Root cause: batch sizing underestimated CJK token density (1.5→1.0 chars/token)
and ignored output budget as primary constraint, causing cascading finish_reason=length.

Changes:
- _token_budget.py: CJK_RATIO 1.5→1.0, OTHER_RATIO 2.2→1.8, safety factors 0.75/0.70
- _token_budget.py: new _compute_max_rows_by_output() — output budget is PRIMARY constraint
- _batch_sizer.py: resolve_provider_config() with DB-level context_window/max_output_tokens
- _batch_sizer.py: INPUT_SAFETY_FACTOR applied, max_rows_by_output used as row cap
- _llm_http.py: log actual usage.prompt_tokens/.completion_tokens from provider
- _llm_call.py: retry only missing rows after finish_reason=length (save partial result)
- models/llm.py + schema: provider-level context_window / max_output_tokens (nullable)
- services/llm_provider.py: get_provider_token_config() helper
- Alembic migration: add columns to llm_providers
- Svelte ProviderConfig: collapsible Advanced: Token Limits section
- 12 new tests (token budget, batch sizer, provider config)
- All 492 tests pass

backend/src/plugins/llm_analysis/models.py

@@ -30,6 +30,14 @@ class LLMProviderConfig(BaseModel):
     is_active: bool = True
     is_multimodal: bool = False
     max_images: int | None = None
+    context_window: int | None = Field(
+        None, ge=1000, le=256000,
+        description="Context window in tokens. Leave blank for auto-detection.",
+    )
+    max_output_tokens: int | None = Field(
+        None, ge=256,
+        description="Max output tokens. Must be less than context_window.",
+    )
 # #endregion LLMProviderConfig
 
 # #region ValidationStatus [TYPE Class]