fix(spur-cli): resolve first allocated node from compressed hostlist

[maybeharshit]

What

srun and sattach connect directly to a job's first allocated node to stream
live output. They extracted that node from JobInfo.nodelist with
nodelist.split(',').next() — but the controller reports nodelist as a
compressed Slurm-style hostlist (built by spur_core::hostlist::compress),
not a raw comma list. For any job on 2+ nodes this yields a non-hostname:

• node[001-002] has no top-level comma, so the split returns the whole literal
  node[001-002].
• node[001,003] has a comma only inside the brackets, so the split returns the
  broken fragment node[001.
• multi-prefix (gpu[001-004],cpu[001-002]) splits into the fragment gpu[001-004.

The result was used directly as a host: sattach dialed
http://node[001-002]:6818 and failed to connect, and srun's get_node()
probe failed so live output was silently skipped. Fixes #575.

Approach

• Add spur_core::hostlist::expand_first, which expands a pattern only far
  enough to yield its first hostname — the inverse of the controller's
  compress, and cheaper than expand(..).next() for large allocations.
• Add a shared spur_cli::nodelist::first_allocated_node helper that both call
  sites now use, replacing the duplicated split(',') logic. It falls back to a
  comma-split for a malformed pattern so a bad input still yields something
  connectable rather than nothing (mirroring expand_hostlist_or_split).
• Wire it into sattach (clear error if unresolvable) and srun
try_stream_output (keeps the existing skip-on-failure behavior).

The unrelated srun::first_node helper (for user-typed -w, resolved
server-side) is intentionally left unchanged.

Design choices

• Uses the exact inverse of what produced the string; [1] PR fix(spur-cli): compress sinfo NODELIST as a Slurm-style hostlist #574
  hardened compress() and verified expand(compress(x)) round-trips, so
  relying on expand/expand_first is guaranteed correct even for mixed
  zero-padding (node[9,010-011]) and multi-prefix lists.
• Client-side only: no proto, persisted Raft/WAL, or config change, so there is
  no upgrade/compat hazard.

Testing

• Unit: new expand_first tests (range, gap, multi-prefix, mixed padding, plain
  list, single, empty, suffix bracket, unmatched-bracket error, and parity with
  expand(..).next()), plus first_allocated_node tests. cargo clippy
  (spur-core, spur-cli) clean; cargo test for spur-core and spur-cli pass.
• E2E on a 2-node LXD cluster with hostnames compressing to spurnode[1-2]:
  before the fix, sattach failed with a DNS error on spurnode[1-2]; after,
  sattach connects to spurnode1 and streams live output, and srun -N2
  streams from both nodes with no controller-side connect errors.

Made with Cursor