Add flag to select tfrt backend for CPU.

[zhangqiaorjc]

The new TFRT CPU backend has been running without issues internally for a while.

Adding an option to use it in open source.

Benchmarks show significant dispatch performance improvements (wall time)

name                                  old time/op             new time/op             delta
eager_unary_dispatch                  41.0µs ± 1%             32.7µs ± 2%  -20.30%         (p=0.000 n=10+9)
eager_unary                           41.8µs ± 1%             33.5µs ± 2%  -19.94%         (p=0.000 n=10+9)
eager_binary_dispatch                 50.4µs ± 1%             41.3µs ± 2%  -17.88%         (p=0.000 n=10+9)
eager_binary                          51.2µs ± 0%             42.0µs ± 2%  -18.02%          (p=0.000 n=9+9)
jit_trivial_dispatch                  55.1µs ± 2%             54.9µs ± 2%     ~            (p=0.604 n=10+9)
jit_trivial                           55.8µs ± 2%             55.7µs ± 2%     ~            (p=0.905 n=10+9)
jit_simple_dispatch                   8.99µs ± 1%             2.13µs ± 4%  -76.26%        (p=0.000 n=10+10)
jit_simple                            10.0µs ± 4%              2.5µs ± 4%  -74.84%          (p=0.000 n=9+8)
jit_simple_many_args_dispatch_10      16.4µs ±11%              4.4µs ±16%  -73.13%        (p=0.000 n=10+10)
jit_simple_many_args_10               17.0µs ±11%              4.6µs ± 2%  -73.10%         (p=0.000 n=10+8)
jit_simple_pruned_args_dispatch_10    8.85µs ± 1%             2.50µs ± 3%  -71.75%          (p=0.000 n=8+8)
jit_simple_pruned_args_10             10.6µs ±12%              2.9µs ± 4%  -72.34%         (p=0.000 n=10+8)
jit_simple_many_args_dispatch_100     75.3µs ±11%             25.7µs ± 2%  -65.82%         (p=0.000 n=10+8)
jit_simple_many_args_100              78.0µs ± 9%             26.0µs ± 3%  -66.61%         (p=0.000 n=10+8)
jit_simple_pruned_args_dispatch_100   16.7µs ± 9%              8.8µs ± 3%  -47.55%         (p=0.000 n=10+8)
jit_simple_pruned_args_100            17.6µs ± 9%              9.7µs ±16%  -44.83%        (p=0.000 n=10+10)
jit_simple_many_args_dispatch_1000     606µs ± 1%              309µs ± 4%  -49.01%         (p=0.000 n=9+10)
jit_simple_many_args_1000              622µs ± 2%              314µs ± 5%  -49.51%        (p=0.000 n=10+10)
jit_simple_pruned_args_dispatch_1000  83.8µs ± 3%             73.7µs ± 4%  -12.10%         (p=0.000 n=10+9)
jit_simple_pruned_args_1000           84.3µs ± 1%             75.7µs ± 1%  -10.17%          (p=0.000 n=9+7)
jit_simple_many_args_dispatch_2000    1.22ms ± 2%             0.69ms ± 5%  -43.39%          (p=0.000 n=9+9)
jit_simple_many_args_2000             1.23ms ± 3%             0.70ms ± 5%  -43.32%          (p=0.000 n=9+9)
jit_simple_pruned_args_dispatch_2000   179µs ± 3%              167µs ± 4%   -7.08%          (p=0.000 n=9+9)
jit_simple_pruned_args_2000            180µs ± 2%              168µs ± 5%   -6.82%          (p=0.000 n=9+9)
jit_dispatch_without_transfer          797µs ± 5%              683µs ± 6%  -14.32%        (p=0.000 n=10+10)
jit_dispatch_with_transfer             817µs ± 8%              759µs ± 5%   -7.07%        (p=0.001 n=10+10)
sda_index_1                           12.6µs ±11%             11.4µs ± 8%   -9.56%         (p=0.022 n=10+9)

[hawkinsp]

Should this registration still be here? I think it's subsumed by the ones above.

[zhangqiaorjc]

removed