Bugfix in running TapedFunction.

[yebai]

I noticed a subtle bug while investigating the performance benchmarks in #133. It turns out that we never run the taped function in @btime (more precisely, tf is only run once) because the tf.counter variable is not reset after each run.

Benchmark results for this PR:

benchmarking rosenbrock...
  Run Original Function:  390.527 μs (16 allocations: 6.10 MiB)
  Run TapedFunction:  479.145 μs (282 allocations: 6.11 MiB)
  Run TapedFunction (compiled):  497.135 μs (309 allocations: 6.11 MiB)
benchmarking ackley...
  Run Original Function:  1.396 ms (0 allocations: 0 bytes)
  Run TapedFunction:  770.484 ms (16698219 allocations: 445.53 MiB)
  Run TapedFunction (compiled):  743.913 ms (18498246 allocations: 592.02 MiB)
benchmarking matrix_test...
  Run Original Function:  95.125 μs (16 allocations: 469.22 KiB)
  Run TapedFunction:  121.585 μs (322 allocations: 477.91 KiB)
  Run TapedFunction (compiled):  122.064 μs (352 allocations: 480.67 KiB)
benchmarking neural_net...
  Run Original Function:  552.198 ns (4 allocations: 576 bytes)
  Run TapedFunction:  4.707 μs (75 allocations: 3.06 KiB)
  Run TapedFunction (compiled):  4.596 μs (83 allocations: 3.73 KiB)

Before this PR (copied from #133 (comment))

benchmarking rosenbrock...
  Run Original Function:  998.907 μs (16 allocations: 6.10 MiB)
  Run TapedFunction:  23.594 ns (0 allocations: 0 bytes)
  Run TapedFunction (compiled):  38.710 ns (1 allocation: 48 bytes)
benchmarking ackley...
  Run Original Function:  2.027 ms (0 allocations: 0 bytes)
  Run TapedFunction:  23.494 ns (0 allocations: 0 bytes)
  Run TapedFunction (compiled):  40.625 ns (1 allocation: 48 bytes)
benchmarking matrix_test...
  Run Original Function:  154.501 μs (16 allocations: [469]
  Run TapedFunction:  22.066 ns (0 allocations: 0 bytes)
  Run TapedFunction (compiled):  38.005 ns (1 allocation: 48 bytes)
benchmarking neural_net...
  Run Original Function:  708.276 ns (4 allocations: 576 bytes)
  Run TapedFunction:  28.743 ns (0 allocations: 0 bytes)
  Run TapedFunction (compiled):  48.381 ns (1 allocation: 48 bytes)

[KDr2]

Hmmm... this result makes much more sense now...

[yebai]

Hmmm... this result makes much more sense now...

@KDr2 It might be worth analysing the new benchmark results to improve the implementation (e.g. type stability, allocation). We can talk a bit more when we meet. I think TapedFunction should target a similar performance of ReverseDiff.Tape's forward pass.