These changes based on `optimized-20170205` (which should've been
`optimized-20170215`) remove the remaining subshells and pipelines that
have a measurable impact without altering functionality. All pipelines
have been refactored away, and the only subshells that remain include:

- The call to `resolve_link` from `abs_dirname` in `libexec/bats`, since
  it needs to invoke one of `greadlink`, `readlink`, or `true`. However,
  it's been refactored to eliminate its own command substitution
  containing a pipeline, and searches for the appropriate system command
  only on the first call.
- The invocation of `tput cols` in `update_screen_width` from
  `libexec/bats-format-tap-stream`.
- The invocation of the command passed as the arguments to `run` from
  `libexec/bats-exec-test`, naturally, since the command must execute in
  its own process and have its output captured.
- The invocation of the command passed as arguments to `buffer` from
  `libexec/bats-format-tap-stream`. It not only avoids an annoying
  flicker in the test output compared to removing it altogether, but
  somehow actually causes this formatting module to perform slightly
  better compared to no buffering at all. My guess as to why it performs
  better is that in this case, making several terminal I/O calls with
  lots of formatting codes may be more expensive than capturing all the
  text and formatting codes in subshells and emitting a single terminal
  I/O call. Also, using a `while read` loop to append to a string didn't
  appear to provide a noticeable performance benefit, so the original
  implementation remains unchanged.

While the performance improvements are not as dramatic as those from
`optimized-20170205`, they are still significant, and only required a
few straightforward changes to the existing code.

All times were collected on a MacBook Pro with a 2.9GHz Intel Core i5
CPU and 8GB 1867MHz DDR3 RAM. The macOS times were collected on macOS
10.12.3. Times for other operating systems were collected on the same
machine, running under VMmare Fusion 8.5.5.

These first two sets of times are for the current Bats test suite on
macOS, which now runs O(1sec) or ~23% faster.

macOS/Bash 3.2.57(1)-release before:
  46 tests, 0 failures

  real    0m4.429s
  user    0m2.627s
  sys     0m1.473s

After:
  real    0m3.404s
  user    0m2.163s
  sys     0m0.850s

macOS/Bash 4.4.12(1)-release before:
  46 tests, 0 failures

  real    0m4.497s
  user    0m2.526s
  sys     0m1.529s

After:
  real    0m3.433s
  user    0m2.038s
  sys     0m0.890s

The remaining times are for the https://github.com/mbland/go-script-bash
test suite at commit c150736fcec0d21a83510d8c6a92e22da733c369 (from
https://github.com/mbland/go-script-bash/pull/166). Improvements are in
the 7%-8% range across the board on most platforms, but as expected, are
significantly greater on some Windows implementations, reaching up to
12%-13.5%.

macOS/Bash 3.2.57(1)-release before:
  789 tests, 0 failures, 2 skipped

  real    1m25.806s
  user    1m1.779s
  sys     0m18.484s

After (~7% faster):
  real    1m19.487s
  user    0m57.632s
  sys     0m15.909s

macOS/Bash 4.4.12(1)-release before:
  789 tests, 0 failures, 3 skipped

  real    1m20.796s
  user    0m55.405s
  sys     0m18.684s

After (~6% faster):
  real    1m14.554s
  user    0m51.306s
  sys     0m16.101s

Ubuntu Linux 16.10 (yakkety)/Bash 4.3.46(1)-release before:
  789 tests, 0 failures, 3 skipped

  real    1m9.426s
  user    0m31.220s
  sys     0m4.224s

After (~8% faster):
  real    1m3.911s
  user    0m29.856s
  sys     0m3.948s

Arch Linux/Bash 4.4.12(1)-release before:
  789 tests, 0 failures, 4 skipped

  real    0m46.714s
  user    0m30.503s
  sys     0m3.713s

After (~9% faster):
  real    0m42.468s
  user    0m28.643s
  sys     0m3.693s

Windows 10/Windows Subsystem for Linux/Bash 4.3.11(1)-release before:
  789 tests, 0 failures, 3 skipped

  real    3m37.525s
  user    0m36.969s
  sys     3m0.891s

After (~7.5% faster):
  real    3m21.062s
  user    0m34.078s
  sys     2m44.313s

Windows 10/Cygwin/Bash 4.4.12(3)-release before:
  789 tests, 0 failures, 3 skipped

  real    5m9.691s
  user    1m33.075s
  sys     2m44.395s

After (~12% faster):
  real    4m32.413s
  user    1m22.036s
  sys     2m22.253s

Windows 10/Git for Windows/Bash 4.3.46(2)-release run from the Windows
Command Prompt (not the Git for Windows MSYS2 terminal) before:
  789 tests, 0 failures, 14 skipped

  real    5m24.973s
  user    1m34.273s
  sys     2m47.360s

After (~13.5% faster):
  real    4m41.114s
  user    1m23.613s
  sys     2m21.771s