feat: support redirects in pipe sequence commands

[bartlomieju]

The parser previously rejected redirects on any command in a pipe
sequence (e.g. cmd 2>&1 | tee log), even though the AST's Command
node already carries a per-command redirect and the executor's
execute_command applies it before dispatching to execute_simple_command
/ execute_subshell. Drop the parser bailout in parse_pipeline_inner
so common bash idioms like cmd 2>&1 | reader and cmd > file | reader
parse and execute.

The existing negative parser test that asserted the error message is
replaced with two positive parser tests covering echo 1 1> stdout.txt | cat and cmd 2>&1 | tee log. Two integration tests exercise the
runtime behavior end-to-end: that the left side's stdout redirect makes
the right side receive an empty pipe, and that 2>&1 | correctly
funnels both stdout and stderr into the pipe.

Fixes denoland/deno#28808.

[CLAassistant]

All committers have signed the CLA.

[bartlomieju]

Reviewed — clean, correct, and well-tested. I traced the executor to confirm the central claim, and the semantics match bash exactly.

The fix is sound. The parser bailout was overly conservative; the executor already handles per-command redirects inside a pipe sequence:

• execute_pipe_sequence assigns each command's stdout to the pipe writer, then calls execute_command.
• execute_command resolves the per-command redirect after the pipe stdout is assigned, then reassigns the fds.

This ordering is exactly bash's model (pipe redirection first, then the command's own redirects), so both cases come out right:

• echo 1 > output.txt | cat — pipe sets echo's stdout to the pipe, then > output.txt overrides fd 1 to the file; the pipe writer drops, so cat gets EOF (len=0). ✓
• ... 2>&1 | reader — resolve_redirect_pipe for IoFile::Fd(1) returns the stdout writer, which is the pipe, so fd 2 → pipe; both streams funnel through (o\ne). ✓

Tests improved. The negative parser test is replaced by two positive parser tests (full AST) plus two deterministic end-to-end integration tests covering the redirect-vs-pipe ordering.

Minor: input/middle-command redirects (a | cmd < input | reader) are now also permitted — correct bash behavior and the executor handles it, though not covered by a test. Non-blocking.

One gate note: CI shows "no checks found" — worth confirming the workflows actually run before merge.

@bartlomieju this is ready to merge (pending CI).