pydantic v1 => v2

[dovinmu]

---

Summary

Migrate all pydantic v1 patterns to v2, required because chromadb depends on pydantic>=2 which conflicts with this project's
pydantic<v2 pin.

What changed

This is a mechanical migration with no behavioral changes:

• pyproject.toml: "pydantic<v2" → "pydantic>=2.0"
• @validator("*") → @model_validator(mode="before"): 6 DB configs (api, aws_opensearch, doris, milvus, oss_opensearch, qdrant_cloud,
  tidb) had per-field empty-string validators; these become a single mode="before" model validator with identical logic
• @root_validator → @model_validator(mode="before"): oss_opensearch engine name mapping
• @validator("field") → @field_validator("field", mode="before"): oss_opensearch quantization type coercion, dataset size
  verification
• .dict() → .model_dump(), .json() → .model_dump_json(): 9 call sites
• PrivateAttr → ClassVar: _size_label dicts on dataset classes (7 sites)
• int | None → int | None = None: ~40 Optional fields that now require explicit defaults in v2
• model_config = ConfigDict(validate_default=True): ensures SecretStr fields with string defaults are coerced at construction time
• MetricType = "cosine" → MetricType = MetricType.COSINE: chroma config (enum, not bare string)
• Lint fixes: import sorting, unused import removal, EM102 compliance, black reformatting

Also incorporates the _httpx_host() fix from main (IPv6 localhost resolution on macOS).

What did NOT change

• No benchmark logic, dataset handling, or result reporting was modified
• No new dependencies
• The endee==0.1.10 pin is preserved with a TODO to check v2 compat

Test plan

• CI passes (lint + unit tests)
• Run a VectorDBBench benchmark to verify no runtime regressions