Add benchmarks

[KDr2]

No description provided.

[yebai]

Result for 318064d

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 A few (quick) observations:

1. The matrix_test is nearly ten times slower than the original function.
2. Compiled function always allocates (one) more and about 2x slower.

EDIT: it seems TapedFuntion is always faster!

EDIT 2: Perhaps the increased CI time of Turing.jl is from 1) the additional step of constructing TapedFunction (i.e. transforming the CodeInfo)... 2) the copy step of TapedFunction is suboptimal 3) integration between TapedTask and TapedFunction is suboptimal