fix(gradlew): collapse consecutive blank lines in build filter output

[luinstra]

Summary

• BuildLineFilter::feed_line was stateless: every line filter_build_line accepted got emitted as-is, including blanks intentionally kept as section separators.
• On real kotlin/JVM builds — which emit groups of w: warnings separated by blank lines — this produced multi-screen runs of blank lines between groups of warnings, drowning the actual signal.
• Make BuildLineFilter stateful with last_emitted_was_blank: bool. Single-blank separators are preserved; runs of 2+ collapse to one. Non-blank lines unaffected.

Note on the "fix" framing: BPE tokenizers fully collapse whitespace runs, so this patch saves ~0 tokens. The win is human readability (tee files, debug sessions, scrollback). Filed as fix because it removes pathological output noise that hides the actual build signal.

Test plan

• cargo fmt --all --check && cargo clippy --all-targets -- -D warnings && cargo test — all pass
• Four new tests use byte-exact assert_eq! via a run_build_filter helper:
  • test_build_filter_collapses_consecutive_blanks — runs of \n\n\n\n collapse to \n\n
  • test_build_filter_preserves_single_blank_separator — single \n\n survives
  • test_build_filter_collapse_survives_dropped_lines_between_blanks — > Task :… stripped between blanks still collapses (regression for the "state only tracks emitted lines" decision)
  • test_build_filter_leading_blanks_collapsed — initial run of blanks collapses to one

Targets develop per CONTRIBUTING.md. Independent of #1939 (same file, unrelated bug).

[YOMXXX]

LGTM as an outside reviewer 👍

The last_emitted_was_blank state-machine pattern is the right shape — preserves a single blank line as a section separator while collapsing the pathological N-blank runs that real kotlin/JVM builds emit between groups of w: warnings. State change is minimal, the flush path is unchanged, and the new tests exercise both the kept-blank and collapsed-blank paths.

Maintainers: clean targeted change. Recommend merging.