* flate: Improve speed in big stateless blocks.Don't re-alloc and copy dict for every block when compressing more than 32KB.```benchmark                         old ns/op     new ns/op     deltaBenchmarkEncodeDigitsSL1e4-32     52954         52850         -0.20%BenchmarkEncodeDigitsSL1e5-32     781061        745420        -4.56%BenchmarkEncodeDigitsSL1e6-32     8143640       7715674       -5.26%BenchmarkEncodeTwainSL1e4-32      68150         68415         +0.39%BenchmarkEncodeTwainSL1e5-32      715140        687326        -3.89%BenchmarkEncodeTwainSL1e6-32      7718175       7339694       -4.90%benchmark                         old MB/s     new MB/s     speedupBenchmarkEncodeDigitsSL1e4-32     188.84       189.21       1.00xBenchmarkEncodeDigitsSL1e5-32     128.03       134.15       1.05xBenchmarkEncodeDigitsSL1e6-32     122.80       129.61       1.06xBenchmarkEncodeTwainSL1e4-32      146.74       146.17       1.00xBenchmarkEncodeTwainSL1e5-32      139.83       145.49       1.04xBenchmarkEncodeTwainSL1e6-32      129.56       136.25       1.05xbenchmark                         old allocs     new allocs     deltaBenchmarkEncodeDigitsSL1e4-32     0              0              +0.00%BenchmarkEncodeDigitsSL1e5-32     3              0              -100.00%BenchmarkEncodeDigitsSL1e6-32     41             0              -100.00%BenchmarkEncodeTwainSL1e4-32      0              0              +0.00%BenchmarkEncodeTwainSL1e5-32      3              0              -100.00%BenchmarkEncodeTwainSL1e6-32      41             0              -100.00%benchmark                         old bytes     new bytes     deltaBenchmarkEncodeDigitsSL1e4-32     0             0             +0.00%BenchmarkEncodeDigitsSL1e5-32     92929         9             -99.99%BenchmarkEncodeDigitsSL1e6-32     1298964       97            -99.99%BenchmarkEncodeTwainSL1e4-32      0             0             +0.00%BenchmarkEncodeTwainSL1e5-32      92928         8             -99.99%BenchmarkEncodeTwainSL1e6-32      1298871       92            -99.99%```* Pin garble to v0.7.2

This patches reduces the number of bounds checks in zstd, as reported bygo build -gcflags="-d=ssa/check_bce/debug=1" |& wc -lfrom 723 to 693.

These are all in return sequences, so no performance gain expected, butthe generated code is smaller.

Also allow seek after EOF.