Prevent the parser from triggering a stack overflow

[lpinca]

This makes the parser work with buffers containing thousands of frames.

Fixes #990.

cc: @Nibbler999

[lpinca]

It seems to be slightly slower on a couple of benchmarks, not sure why.

Before:

ping message x 1,447,412 ops/sec ±1.53% (83 runs sampled)
ping with no data x 14,451,215 ops/sec ±0.86% (84 runs sampled)
close message x 12,488,814 ops/sec ±0.91% (87 runs sampled)
masked text message (20 bytes) x 886,571 ops/sec ±1.71% (90 runs sampled)
binary data (125 bytes) x 1,263,534 ops/sec ±1.06% (89 runs sampled)
binary data (65535 bytes) x 289,829 ops/sec ±1.20% (88 runs sampled)
binary data (200 KiB) x 111,717 ops/sec ±0.79% (90 runs sampled)
binary data (1 MiB) x 18,408 ops/sec ±1.45% (86 runs sampled)

After:

ping message x 1,430,195 ops/sec ±0.77% (86 runs sampled)
ping with no data x 11,501,751 ops/sec ±1.03% (88 runs sampled)
close message x 10,360,093 ops/sec ±0.80% (89 runs sampled)
masked text message (20 bytes) x 881,880 ops/sec ±1.24% (88 runs sampled)
binary data (125 bytes) x 1,274,296 ops/sec ±1.05% (89 runs sampled)
binary data (65535 bytes) x 281,736 ops/sec ±0.93% (91 runs sampled)
binary data (200 KiB) x 110,382 ops/sec ±0.83% (88 runs sampled)
binary data (1 MiB) x 19,236 ops/sec ±0.33% (91 runs sampled)