Migrate to csv crate rewrite.

This commit resists the urge to refactor/rewrite xsv and ports it over
to the new CSV API. It made a lot of things cleaner and even improved
the performance of core commands like `count`, `sample`, `search`,
`select` and `slice`.

This also removes the last remaining (dubious) uses of `unsafe` within
xsv.

Benchmarks before/after:

    benchmark               before                      after
    count                   0.26s   175.05  MB/sec      0.11   413.76  MB/sec
    flatten                 4.53s   10.04   MB/sec      4.54   10.02   MB/sec
    flatten_condensed       4.72s   9.64    MB/sec      4.45   10.22   MB/sec
    frequency               1.91s   23.82   MB/sec      1.82   25.00   MB/sec
    index                   0.28s   162.54  MB/sec      0.12   379.28  MB/sec
    sample_10               0.43s   105.84  MB/sec      0.18   252.85  MB/sec
    sample_1000             0.44s   103.44  MB/sec      0.18   252.85  MB/sec
    sample_100000           0.50s   91.02   MB/sec      0.29   156.94  MB/sec
    search                  0.59s   77.14   MB/sec      0.27   168.56  MB/sec
    select                  0.41s   111.00  MB/sec      0.14   325.09  MB/sec
    sort                    2.59s   17.57   MB/sec      2.18   20.87   MB/sec
    slice_one_middle        0.22s   206.88  MB/sec      0.08   568.92  MB/sec
    slice_one_middle_index  0.01s   4551.36 MB/sec      0.01   4551.36 MB/sec
    stats                   1.26s   36.12   MB/sec      1.09   41.75   MB/sec
    stats_index             0.19s   239.54  MB/sec      0.15   303.42  MB/sec
    stats_everything        2.13s   21.36   MB/sec      1.94   23.46   MB/sec
    stats_everything_index  1.00s   45.51   MB/sec      0.93   48.93   MB/sec

.travis.yml

@@ -1,17 +1,3 @@
-#language: rust
-#rust:
-#  - 1.9.0
-#  - stable
-#  - beta
-#  - nightly
-#script:
-#  - cargo build --verbose
-#  - cargo doc
-#  - cargo test --verbose
-#  - if [ "$TRAVIS_RUST_VERSION" = "nightly" ]; then
-#      cargo bench --verbose;
-#    fi
-
 language: rust
 cache: cargo
 
@@ -33,6 +19,13 @@ matrix:
     - os: linux
       rust: stable
       env: TARGET=x86_64-unknown-linux-musl
+    # Minimum Rust supported channel.
+    - os: linux
+      rust: 1.15.0
+      env: TARGET=x86_64-unknown-linux-gnu
+    - os: linux
+      rust: 1.15.0
+      env: TARGET=x86_64-unknown-linux-musl
 
 before_install:
   - export PATH="$PATH:$HOME/.cargo/bin"

BENCHMARKS.md

@@ -6,27 +6,27 @@ These benchmarks were run with
 which is a random 1,000,000 row subset of the world city population dataset
 from the [Data Science Toolkit](https://github.com/petewarden/dstkdata).
 
-These benchmarks were run on an Intel i3930K (6 CPUs, 12 threads) with 32GB of
-memory.
+These benchmarks were run on an Intel i7-6900K (8 CPUs, 16 threads) with 64GB
+of memory.
 
 ```
-count                   0.28 seconds    162.54 MB/sec
-flatten                 5.31 seconds    8.57 MB/sec
-flatten_condensed       5.39 seconds    8.44 MB/sec
-frequency               2.54 seconds    17.91 MB/sec
-index                   0.27 seconds    168.56 MB/sec
-sample_10               0.47 seconds    96.83 MB/sec
-sample_1000             0.49 seconds    92.88 MB/sec
-sample_100000           0.62 seconds    73.40 MB/sec
-search                  0.71 seconds    64.10 MB/sec
-select                  0.47 seconds    96.83 MB/sec
-sort                    3.36 seconds    13.54 MB/sec
-slice_one_middle        0.22 seconds    206.88 MB/sec
-slice_one_middle_index  0.01 seconds    4551.36 MB/sec
-stats                   1.37 seconds    33.22 MB/sec
-stats_index             0.23 seconds    197.88 MB/sec
-stats_everything        3.90 seconds    11.67 MB/sec
-stats_everything_index  2.58 seconds    17.64 MB/sec
+count                   0.11  seconds  413.76   MB/sec
+flatten                 4.54  seconds  10.02    MB/sec
+flatten_condensed       4.45  seconds  10.22    MB/sec
+frequency               1.82  seconds  25.00    MB/sec
+index                   0.12  seconds  379.28   MB/sec
+sample_10               0.18  seconds  252.85   MB/sec
+sample_1000             0.18  seconds  252.85   MB/sec
+sample_100000           0.29  seconds  156.94   MB/sec
+search                  0.27  seconds  168.56   MB/sec
+select                  0.14  seconds  325.09   MB/sec
+sort                    2.18  seconds  20.87    MB/sec
+slice_one_middle        0.08  seconds  568.92   MB/sec
+slice_one_middle_index  0.01  seconds  4551.36  MB/sec
+stats                   1.09  seconds  41.75    MB/sec
+stats_index             0.15  seconds  303.42   MB/sec
+stats_everything        1.94  seconds  23.46    MB/sec
+stats_everything_index  0.93  seconds  48.93    MB/sec
 ```
 
 ### Details
@@ -39,4 +39,3 @@ The `count` command can be viewed as a sort of baseline of the fastest possible
 command that parses every record in CSV data.
 
 The benchmarks that end with `_index` are run with indexing enabled.
-