Add schema inference from real sample data (tb datasource analyze parity)

[sandshoes]

Summary

• TinybirdApi.analyze() wraps the Tinybird Analyze API (POST /v0/analyze), inferring column names, ClickHouse types, and (for CSV) dialect from a local file or URL — matching tb datasource analyze <url_or_file>. Follows the same url/file (form-encoded vs. multipart) convention as the existing append_datasource().
• tinybird_sdk.infer.parse_analyze_response() parses the raw API response into a structured AnalyzedSchema, tolerating two known response shapes: an analysis.columns list of {name, recommended_type} entries, or an analysis.schema/schema ClickHouse DDL-like string as a fallback.
• tinybird_sdk.infer.generate_schema_code() turns an AnalyzedSchema into ready-to-paste define_datasource(...) source with t.* validators, reusing the existing codegen.clickhouse_type_to_validator type-mapping (the same one used for reverse codegen of live workspace resources), so inferred and reverse-codegen'd datasources read identically.

Closes https://linear.app/tinybird/issue/PROTM-2263/python-sdk-add-schema-inference-from-real-sample-data

Design notes

• This PR does not add a new tinybird analyze CLI subcommand — it ships the library-level helper (API method + parsing + codegen) the ticket asks for. A thin CLI wrapper mirroring tb datasource analyze would be a natural, separable follow-up once this API shape is validated against the real Analyze API contract in a live environment.
• The exact response shape of /v0/analyze could not be verified against live docs/API from this sandbox, so parse_analyze_response() is intentionally tolerant of both a columns list and a DDL-style schema string, and raises a clear ValueError if neither is present rather than silently guessing. Worth a quick confirmation against a real analyze response before relying on this in production.

Checklist

• CI is green (lint, typecheck, test)
• pre-commit run --all-files passes locally (secrets/gitleaks fails in this sandbox on an unrelated SSL cert error fetching its pre-commit env, not something introduced by this change)
• Tests were added or updated when behavior changed
• Public API / typing changes were reviewed
• Documentation was updated (README.md) — new "Schema Inference Helpers" example
• Breaking changes are clearly documented (none — purely additive)
• CHANGELOG.md was updated when user-facing behavior changed