## Description

- Add _estimate_tokens, _chunk_text, _chunk_and_embed to EmbedderBase
- Override embed()/embed_batch() in OpenAIDenseEmbedder with auto-chunking
- Use tiktoken for accurate token counting, fallback to char estimation
- Merge chunk vectors via token-weighted average + L2 normalization
- Make max_tokens configurable via ov.conf (embedding.dense.max_tokens)
- Default 8000 for OpenAI models, users can override for other models
- Add 22 unit tests covering chunking, fallback, batch, config

## Related Issue

https://github.com/volcengine/OpenViking/issues/616
Fixes #616

## Type of Change

- [x] Bug fix (non-breaking change that fixes an issue)
- [ ] New feature (non-breaking change that adds functionality)
- [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
- [ ] Documentation update
- [x] Refactoring (no functional changes)
- [x] Test update

## Changes Made

- `openviking/models/embedder/base.py`: Added `max_tokens` property, `_estimate_tokens()`, `_chunk_text()` (paragraph > sentence > fixed-length splitting), `_merge_segments()`, `_fixed_length_split()`, `_chunk_and_embed()` (token-weighted average + L2 normalization), and `_embed_single()` to EmbedderBase
- `openviking/models/embedder/openai_embedders.py`: Refactored `embed()` → `_embed_single()`, added auto-chunking in `embed()` and `embed_batch()`, initialized tiktoken encoder at construction time, made `max_tokens` configurable via constructor parameter
- `openviking_cli/utils/config/embedding_config.py`: Added `max_tokens` field to `EmbeddingModelConfig`, passed through to OpenAIDenseEmbedder in factory registry
- `tests/unit/test_openai_embedder_chunking.py`: 22 unit tests covering short text (no chunking), oversized text, very long text (100x limit), empty text, batch with mixed lengths, L2 normalization, paragraph/sentence/fixed-length splitting, tiktoken fallback, and configurable max_tokens

## Testing

- [x] I have added tests that prove my fix is effective or that my feature works
- [x] New and existing unit tests pass locally with my changes
- [x] I have tested this on the following platforms:
  - [x] Linux
  - [ ] macOS
  - [ ] Windows

## Checklist

- [x] My code follows the project's coding style
- [x] I have performed a self-review of my code
- [x] I have commented my code, particularly in hard-to-understand areas
- [x] I have made corresponding changes to the documentation
- [x] My changes generate no new warnings
- [ ] Any dependent changes have been merged and published

## Screenshots (if applicable)

N/A

## Additional Notes

- `tiktoken` is an optional dependency: when unavailable (e.g., non-OpenAI models via compatible API), token estimation falls back to `len(text) // 3`
- `max_tokens` can be configured in `ov.conf` under `embedding.dense.max_tokens`, allowing users to set appropriate limits for different models (e.g., 32000 for Qwen3-Embedding-4B, 8000 default for OpenAI models)
- The chunking logic is implemented at the `EmbedderBase` level so other embedders can benefit in the future, but only `OpenAIDenseEmbedder` activates it in this PR
- `embed_batch()` intelligently routes: short texts are batched via the native OpenAI batch API for efficiency, while only oversized texts are individually chunked
- Verified locally with `max_tokens=100` against SiliconFlow API (Qwen3-Embedding-4B): no 400 errors, semantic artifacts generated correctly