perf(source-map): reuse resolved segments in range queries

[luojiyin1987]

What

Reuse the already-resolved segment indices in resolveSegmentRange() instead of calling sourceOffsetAt() (which re-does the same binary search via findSegmentAt()).

Why

A non-empty range query performed 4 x O(log segments) binary searches:

1. findSegmentIndexAt(segments, valueStart) — kept
2. findSegmentIndexAt(segments, valueEnd - 1) — kept
3. sourceOffsetAt() -> findSegmentAt(segments, valueStart) — removed
4. sourceOffsetAt() -> findSegmentAt(segments, valueEnd - 1) — removed

Steps 3-4 are now inlined using the already-resolved indices from steps 1-2.

Changes

• resolveSegmentRange(): inline offset calculation from the resolved segment indices; remove valueLength from destructuring (no longer needed).
• sourceOffsetAt(): removed (zero callers after inline).
• bench-source-map.mjs: add random range queries (10k random range queries per size), rename random queries to random single-unit queries, increase query full value range from 1k to 10k iterations, switch query patterns from time() to timeMedian() for stable A/B comparison.

Benchmark (A/B, 3-sample median)

random single-unit queries          baseline (4x)  optimized (2x)  speedup
1 KiB                                    0.10 ms       0.10 ms        —
16 KiB                                   2.51 ms       0.89 ms       ~65%
64 KiB                                   8.29 ms       4.65 ms       ~44%
256 KiB                                 43.45 ms      30.18 ms       ~31%

random range queries (10k iterations)    baseline (4x)  optimized (2x)  speedup
1 KiB                                    2.97 ms       1.94 ms       ~35%
16 KiB                                   4.55 ms       2.25 ms       ~51%
64 KiB                                   4.29 ms       4.18 ms        ~3%
256 KiB                                  6.23 ms       4.58 ms       ~26%

query full value range (10k iterations)  baseline (4x)  optimized (2x)  speedup
16 KiB                                   1.41 ms       0.89 ms       ~37%
64 KiB                                   1.98 ms       0.92 ms       ~54%
256 KiB                                  1.39 ms       1.08 ms       ~22%

random single-unit queries consistently shows 31-65% speedup across sizes. random range queries is noisier because the optimization only eliminates the second pair of searches — the first pair (already present before) dominates the variance.

Verification

• 1961/1961 tests pass
• Bundle: 186.0 kb (down from 186.1 kb — dead code removal)