feat: Upgrade timestamp precision from milliseconds to microseconds

[xe-nvdk]

Improves timestamp storage from millisecond (ms) to microsecond (us) precision across Arc for better observability and distributed tracing support.

• Sub-millisecond precision: Critical for distributed tracing spans

• Industry standard: Microseconds is standard for observability databases (TimescaleDB, InfluxDB traces, OpenTelemetry)

• No storage overhead: Both ms and us use 64-bit timestamps

• DuckDB native: No conversion needed - reads timestamp[us] natively

• Better precision: 0.001ms resolution vs 1ms resolution

• Schema inference now uses pa.timestamp('us') instead of pa.timestamp('ms')

• Added intelligent timestamp unit auto-detection based on magnitude:

  • < 1e10: seconds (e.g., 1730246400)
  • < 1e13: milliseconds (e.g., 1730246400000) ← Telegraf, load tests
  • ≥ 1e13: microseconds (e.g., 1730246400000000)

• Full backwards compatibility with existing millisecond timestamps

• Added auto-detection for columnar and single-record formats

• Maintains backwards compatibility with msgpack API (still accepts ms)

• Automatically converts to datetime for Arrow writer

• Updated to output datetime objects instead of converting to milliseconds

• Simplified timestamp handling (arrow_writer now accepts datetime directly)

• Supports seconds, milliseconds, and microseconds in query results

NO BREAKING CHANGES - All existing clients continue to work:

• ✅ Telegraf Arc plugin: Sends UnixMilli() → auto-detected as ms

• ✅ Arc telemetry server: Sends milliseconds → auto-detected as ms

• ✅ Load test scripts: Send ms timestamps → auto-detected as ms

• ✅ Line protocol API: Has precision parameter (ns/us/ms/s)

• ✅ Existing data: Old ms data still readable by DuckDB

• Verified new Parquet files store timestamp[us]

• Confirmed DuckDB reads timestamp[us] natively (no conversion)

• Tested JSON endpoint: Returns correct ISO timestamps

• Tested Arrow endpoint: Returns timestamp[us] with full precision

• Verified Telegraf compatibility with millisecond timestamps

• Confirmed mixed precision queries work (old ms + new us data)

No migration needed! New data uses microseconds, old data (ms) remains readable. DuckDB automatically handles mixed precision when querying.