feat: summarization chunking

[Ada-lave]

Summary

Long recordings (1h+) can produce 50–200K characters of transcript text, which exceeds the context window of most local models (e.g. Ollama llama3.2 with 8K ctx ≈ ~10K chars for Russian). Previously the entire transcript was sent in a single LLM request with no length checking, causing silent failures or truncated summaries.

This PR implements a Map-Reduce summarization strategy:

• If the transcript fits within the configured limit → single request, zero change to existing behavior
• If the transcript is too long → split into segment-boundary chunks, summarize each sequentially with rolling context (each chunk receives the previous partial summary as context), then synthesize all partial summaries into a single coherent result

Changes

• src/lib/llm/chunking.ts (new) — core chunking logic: segment-boundary splitting, rolling-context chunk prompts, synthesis prompt, summarizeWithChunking() function with optional onProgress callback
• src/lib/llm/index.ts — added maxInputChars?: number field to LlmConfig
• src/lib/config.ts — added llmDefaultMaxInputChars = 24_000 (≈ 6000 tokens, safe for small local models)
• src/pages/home/view-model.ts — replaced both llm.ask() call sites (auto-summarize + re-summarize) with summarizeWithChunking(); toast now shows live progress: "Summarizing part 2 of 4..." → "Merging summaries..."
• src/pages/batch/view-model.tsx — same replacement for batch mode
• src/components/params.tsx — added Max Input Characters input in LLM settings
• locales/en-US/common.json, locales/ru-RU/common.json — added translation keys for the new setting and progress messages

How it works

Transcript (100K chars)
    ├─ ≤ 24K → single request  [unchanged behavior]
    └─ > 24K → chunk by segment boundaries
          ├─ Chunk 1 ─────────────────────► Partial 1
          ├─ Chunk 2 + [ctx: Partial 1] ──► Partial 2
          ├─ Chunk 3 + [ctx: Partial 2] ──► Partial 3
          └─ Synthesize [P1 + P2 + P3] ──► Final summary

Configuration

Default maxInputChars = 24000. For models with 8K context window, recommended values: 10000–12000 for Russian, 16000–18000 for English.

Closes #999