feat(CLI): add --parallel flag to phrase pull for concurrent downloads

[Thibaut Etienne (tetienne)]

We run phrase pull in CI pipelines with 379 locale files across 3 projects. Sequential downloads take over a minute per run, and the latency adds up across branches and environments. This PR adds parallel downloads to dramatically reduce pull time.

Pairs well with #1066 (--cache for ETag-based conditional requests): parallel mode makes the initial pull fast, and caching makes subsequent pulls nearly free by returning 304 Not Modified (which don't count against rate limits).

Summary

• Add --parallel flag (-p) to phrase pull that downloads locale files using up to 4 concurrent requests (matching the Phrase API concurrency limit)
• Uses errgroup with SetLimit(4) for bounded parallelism
• Results are collected and printed in original locale order after all downloads complete (clean, deterministic output)
• A shared mutex coordinates rate-limit pauses across all workers
• Only supported in sync mode; --parallel with --async warns and ignores
• Extracts buildDownloadOpts helper to eliminate code duplication between sequential and parallel paths

Usage

# Parallel download (4 concurrent requests)
phrase pull --parallel

# Short flag
phrase pull -p

# Without flag: sequential as before (no regression)
phrase pull

Real-world results

Tested against a production project with 379 locale files across 3 Phrase projects:

$ time phrase pull
Downloaded en to locales/en/tradfile.json
Downloaded fr to locales/fr/tradfile.json
...
(379 files downloaded sequentially in ~1m 10s)

$ time phrase pull --parallel
Downloaded en to locales/en/tradfile.json
Downloaded fr to locales/fr/tradfile.json
...
(379 files downloaded in parallel in ~11s)

Mode	Time	Speedup
phrase pull	1m 10s	baseline
phrase pull --parallel	11.5s	~6x faster

Design decisions

• 4 concurrent requests: Matches the Phrase API's documented concurrency limit. Going higher would result in rejected requests.
• errgroup with SetLimit: Standard Go concurrency pattern. Provides bounded parallelism, fail-fast error propagation via context cancellation, and clean Wait() semantics.
• Ordered output: Results are stored in a pre-allocated slice indexed by position. Output is printed after all downloads complete, preserving the original locale file order regardless of goroutine completion order.
• Rate-limit mutex: When any worker hits HTTP 429, it holds a shared mutex while waiting for the rate limit window to reset. Other workers block on the mutex before making their next request, preventing a thundering herd.
• Shared buildDownloadOpts: Extracted the download options preparation (params, file format, tags) into a helper used by both sequential and parallel paths, eliminating ~20 lines of duplication.
• Timeout: Parallel path uses context.WithTimeout matching the existing 30-minute timeout from the sequential path.

Test plan

• go build ./cmd/internal/... compiles
• go test ./cmd/internal/... passes
• go vet ./cmd/internal/ clean
• Manual test: phrase pull --parallel downloads 379 files in 11.5s
• Manual test: phrase pull without --parallel behaves identically
• phrase pull --parallel --async warns and ignores parallel