Fix dataset export fidelity and atomic upload failures

[Gautam Sharma (GautamSharma99)]

Summary

Closes #226

Dataset export/upload now has a versioned, loss-aware round-trip format and validates the complete import before mutating LangSmith.

What changed

• Export uses version 1 of a JSON envelope and preserves dataset name/description plus example IDs, creation timestamps, inputs, outputs, metadata, native split membership, and attachment URLs.
• Upload validates every array member before creating the destination dataset. Non-object members include their array index in the error, and malformed inputs, IDs, timestamps, split values, and unsupported attachment restoration fail early.
• Upload uses the generated SDK bulk-example endpoint so metadata, IDs, timestamps, and splits are sent as native API fields instead of being dropped or encoded in metadata.
• If bulk creation fails after the dataset is created, upload deletes that dataset. If cleanup also fails, the returned error includes both failures so the operator has an actionable dataset ID.
• Legacy array/object input remains accepted, but now requires an explicit inputs object rather than silently treating arbitrary fields as inputs.
• The README documents the versioned schema and attachment behavior.

Tests

• Reject mixed arrays with an indexed validation error.
• Verify metadata and multi-split values survive parsing.
• Reject attachment-bearing imports before remote mutation.
• go test ./...
• go test -race ./internal/cmd
• go vet ./...
• make build