perf(orchestrator): coalesce NBD header + max payload in one read

[AdaAibaby]

fix #3355

Problem

dispatchBufferSize was exactly 4 MiB — the maximum typical data payload size. Every NBD request also carries a 28-byte header, so a max-size (4 MiB) write required two Read syscalls instead of one:

First Read:  [28B header] [4MiB - 28B payload]   ← fills 4 MiB buffer exactly
Second Read: [remaining 28B of payload]

The code had a TODO acknowledging this:

// TODO: Look into optimizing the buffer reads by increasing the buffer size by 28 bytes,
// to account for a request that is 28 bytes of header + 4MB of data.
dispatchBufferSize = 4 * 1024 * 1024

Fix

• Add nbdRequestHeaderSize = 28 named constant — replaces all bare 28 literals in the dispatch loop, making the header-size relationship explicit.
• Increase dispatchBufferSize to 4*1024*1024 + nbdRequestHeaderSize.
• Remove the resolved TODO.
• Simplify the now-inaccurate inline comment in the write path.

With the new buffer size, [28B header + 4MiB payload] fits in a single Read call, saving one syscall per max-size request. For I/O-heavy sandboxes doing large sequential block writes, this adds up across thousands of requests.

Memory impact: 28 bytes per sandbox connection — negligible at any scale.

Tests

• TestDispatchMaxSizeWrite — sends a 4 MiB write (header + full payload as one chunk) and verifies the payload reaches the backend with correct offset and full data integrity.
• nbdRequest() helper updated to use nbdRequestHeaderSize instead of a bare 28.

/cc @jakubno @dobrac @ValentaTomas @arkamar @tvi Looking forward to your code review.