Speed up SMB shared-folder transfers

[felixrieseberg]

SMB copies were much slower than ISO copies. Three independent causes:

1. Stop-and-wait TCP. v86's userspace TCP sends one 1460-byte segment then blocks on pending=true until the guest ACKs. Combined with Win95's delayed-ACK, a 64K READ_RAW becomes ~45 emulated round-trips. setupSmbShare now overrides pump() per-connection to keep a window of segments in flight (up to the guest's advertised rwin, capped at 8×MSS so we never overrun the NE2000 RX ring). The link is in-process and lossless — and the original stack has no retransmit anyway — so this is safe.

2. Per-byte reply assembly. buildSmb() and coreRead() pushed every payload byte into a JS number[] then re-copied — ~32K Array.push calls per 16K read. Both now allocate one Uint8Array and set() the payload.

3. Hot-path diagnostics. The net0-send frame counter and the console.log→appendFileSync tee ran unconditionally on every frame. Both are now gated behind WIN95_SMB_LOG.

ISO will always win (ATAPI is readSync → guest RAM with no protocol/NIC layer), but the windowing should make SMB severalfold faster.

Typecheck clean; all 57 SMB protocol tests pass. The new buildSmb is byte-identical to the old Writer-based one (verified by diff + parseSmb round-trip).