Use NIO for file decryption

[ehrmann]

Refactor the decryption code to use NIO classes. There's a pretty good performance win, and I think it's due to a combination of larger buffers and fewer intermediate buffers.

I tested it by extracting with the modified code and comparing the restored files.

To get an estimate of the performance win, I wrote quick test script that uses similar decrypt implementations over different file sizes:

https://gist.github.com/ehrmann/e80e4eda48b839d41d9cd52d0ba551ee

Windows:

 size,       old,       new
   1K,      1969,      1399
   2K,      2066,      1339
   4K,      1893,      1357
   8K,      1921,      1355
  16K,      2172,      1390
  32K,      2270,      1484
  64K,      2632,      1518
 128K,      2767,      1684
 256K,      3469,      1962
 512K,      5074,      2643
1024K,      7862,      3724
2048K,     13456,      6070
4096K,     26723,     12902

FreeBSD

 size,       old,       new
   1K,        60,        42
   2K,        96,        21
   4K,        17,        42
   8K,        34,        28
  16K,        44,        37
  32K,        83,        55
  64K,       120,        78
 128K,       189,       135
 256K,       397,       144
 512K,       306,       376
1024K,      1136,       591
2048K,      1444,       640
4096K,      3563,      2326

Linux

 size,       old,       new
   1K,       105,        35
   2K,       107,        34
   4K,       107,        32
   8K,       120,        42
  16K,       127,        40
  32K,       165,        47
  64K,       192,        64
 128K,       278,        98
 256K,       438,       160
 512K,       773,       292
1024K,      1422,       525
2048K,      2680,      1047
4096K,      5207,      2020