feat(query): implement a few missing and useless options + fix query show

[redox]

Extend query with lakehouse QueryRequest / output options aligned to the OpenAPI spec, resulting in things like:

altertable query "select * from events LIMIT 10" --catalog whatever --schema main --format parquet > /tmp/events.parquet

New query flags

• --compute-size (XS|S|M|L|XL|AUTO, default AUTO; omitted with --session-id unless an explicit size is set; reject AUTO + --session-id)
• --dialect, --catalog, --schema (sent only when set)
• --format csv|jsonl|parquet|markdown — csv/jsonl/parquet request API format and stream response bytes; markdown stays CLI-rendered from NDJSON
• --output <path> — write result to a file (stdout by default)

Validation & output

• Reject API-native formats with --json/--agent or presentation flags (--layout, --columns, --max-width, --pager)
• Add executeLakehouseQueryBytes and writeQueryDestination for binary-safe streaming
• Add OutputSink.writeBytes so parquet/csv/jsonl redirects are not corrupted by console.log (UTF-8 decode + trailing newline)

Also, fix query show so it always requires a query id and hits GET /query/:query_id.

query show / query cancel

• Remove soleDirectOperands: ["show"], so bare altertable query show no longer runs SQL SHOW (which dumped ducklake metadata).
• query show <query-id> → GET /query/:query_id; query cancel <query-id> --session-id … → DELETE /query/:query_id (unchanged).

[albert20260301]

Required

1. Please reconcile the request contract with the tracked Lakehouse spec before merge. specs/lakehouse/SPEC.md lists the optional QueryRequest fields and does not define dialect or format. Those fields underpin the new flags and raw-response behavior, while the PR body says they are OpenAPI-aligned. Update the spec (or link the spec change) to define the request/response contract, or narrow this PR to the documented fields.

2. writeQueryDestination currently consumes the whole ReadableStream with new Response(stream).arrayBuffer() before writing it. That is binary-safe, but it is not streaming and can exhaust memory for the Parquet/large-result use case this PR introduces. Please pipe chunks to the file/stdout sink instead, with a multi-chunk regression test.

[redox]

Required

1. Please reconcile the request contract with the tracked Lakehouse spec before merge. specs/lakehouse/SPEC.md lists the optional QueryRequest fields and does not define dialect or format. Those fields underpin the new flags and raw-response behavior, while the PR body says they are OpenAPI-aligned. Update the spec (or link the spec change) to define the request/response contract, or narrow this PR to the documented fields.

@albert20260301 good idea, go open a PR on https://github.com/altertable-ai/altertable-client-specs to reflect our updated https://api.altertable.ai/openapi/lakehouse.json

[albert20260301]

@redox I opened altertable-ai/altertable-client-specs#39 from the current live Lakehouse OpenAPI contract, covering the dialect and format fields and removing stale visible.

The binary-output implementation still needs to write the response stream incrementally (with a multi-chunk regression test) before this CLI PR is ready for re-review.