Feature/snowflake benchmarking

[ness-david-dedu]

• Adds a benchmark suite for both Snowflake write connectors under internal/impl/snowflake/bench/ — Taskfiles, setup
  SQL, and configs for snowflake_streaming (Snowpipe Streaming API) and snowflake_put (staged PUT + Snowpipe). Includes
  both one-dimension-at-a-time sweeps and full cross-product matrix tasks (bench:streaming:matrix:full,
  bench:bulk:matrix) for exploring BATCH/PARALLELISM/MAX_IN_FLIGHT/CHUNK_SIZE/UPLOAD_THREADS combinations.
  • Publishes results to docs/benchmark-results/snowflake.md (merged streaming + bulk into one file, two sections,
    condensed with collapsible raw sweep data), linked from docs/benchmarking.md and SUMMARY.md.
  • Fixes a concurrency bug in snowflake_put found while benchmarking: WriteBatch uploaded per-file PUTs serially and
    accessed s.db without synchronization outside Connect/Close.
    • connMut is now an RWMutex guarding only the *sql.DB pointer swap, not held across the whole batch.
    • Files in a batch upload concurrently via errgroup, bounded by upload_parallel_threads.
    • This raised bulk throughput from 4,661 msg/sec (serialized) to 36,166 msg/sec best-found config (BATCH=20000
      MAX_IN_FLIGHT=64 UPLOAD_THREADS=32), closing most of the gap to snowflake_streaming's 40,442 msg/sec ceiling.