perf: use a Map for tracking chunks

[43081j]

Using a map for tracking chunks is faster than a plain object. This is most likely due to v8 thinking we're allocating an array-like object, but then demoting it to a dictionary as we add more keys. That cost of changing allocation will slow it down.

These are @internal properties so it should be fine to change them without a breaking change.

Benchmark results

This branch:

construct x 241,484 ops/sec ±0.31% (102 runs sampled)
append x 20,834,409 ops/sec ±3.14% (70 runs sampled)
indent x 14,808 ops/sec ±1.35% (88 runs sampled)
generateMap (no edit) x 2,971 ops/sec ±1.23% (99 runs sampled)
generateMap (edit) x 3,104 ops/sec ±0.36% (97 runs sampled)
generateDecodedMap (no edit) x 3,263 ops/sec ±2.09% (99 runs sampled)
generateDecodedMap (edit) x 3,353 ops/sec ±0.10% (100 runs sampled)
overwrite x 3.21 ops/sec ±5.65% (12 runs sampled)

Master:

construct x 217,230 ops/sec ±0.51% (97 runs sampled)
append x 20,063,587 ops/sec ±5.06% (68 runs sampled)
indent x 13,950 ops/sec ±1.92% (87 runs sampled)
generateMap (no edit) x 2,994 ops/sec ±1.26% (99 runs sampled)
generateMap (edit) x 3,132 ops/sec ±0.33% (97 runs sampled)
generateDecodedMap (no edit) x 3,328 ops/sec ±0.07% (101 runs sampled)
generateDecodedMap (edit) x 3,357 ops/sec ±0.66% (101 runs sampled)
overwrite x 6.04 ops/sec ±2.08% (20 runs sampled)

its not a huge bump, mostly in construct, but still a bump. i noticed generateMap also speeds up in some cases thanks to this