feat: direct upload with GC-aware addPieces batching

[SgtPooki]

What changed

New upload command implementing #70: download source CIDs, pack multi-root CARs (default 1000 MiB, under the SDK's 1016 MiB piece cap), stream each straight to the primary provider via synapse-sdk store(), have secondaries pull from the primary, and flush parked pieces through one commit() per provider when the batch reaches 40, when the GC-window guess nears expiry, or when the source drains. The per-provider window estimate persists in the migration DB, is lowered on a detected GC, and never raised; the flush margin comes from commit confirmations observed in the same run. No base-fee gate — the provider submits and pays for addPieces. Staged CARs are evicted once every copy is committed.

The provider-pull ingress paths (redirect-serve, serve --ingress, pdp-submit) stay available for self-hosting operators but now require --legacy-pull.

How to verify

Calibnet runs against providers 4 and 2: a 45-piece run flushed [batch-full] at exactly 40 then [drained] the remaining 5 on both providers; two 999.99 MiB pieces stored, pulled, and committed; CARs evicted only after both copies landed. A real GC collection cannot be forced on demand, so detection and window-lowering are covered by gc-window.test.ts and direct-upload-flow.test.ts instead.