PERF: better dtype inference for perf gains (GH7332)

[jreback]

closes #7332

for some reason int64 was not being looked up correctly in the _TYPE_MAP table.
changing to work on name/kind makes this more efficient.

separately better inference of timedelta conversions from i8 were not being done
so that yields a big boost

-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------
dtype_infer_timedelta64_2                    |   3.9814 | 1765.3640 |   0.0023 |
dtype_infer_datetime64                       |   4.8450 | 1825.3207 |   0.0027 |
dtype_infer_int64                            |   0.7904 |  18.5820 |   0.0425 |
dtype_infer_float64                          |   0.8090 |   1.0493 |   0.7710 |
dtype_infer_float32                          |   0.3430 |   0.3823 |   0.8971 |
dtype_infer_int32                            |   0.3897 |   0.4017 |   0.9701 |
dtype_infer_uint32                           |   0.5450 |   0.5583 |   0.9762 |
dtype_infer_timedelta64_1                    | 118.2277 | 117.6037 |   1.0053 |
-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------

Ratio < 1.0 means the target commit is faster then the baseline.
Seed used: 1234

Target [e2f5a50] : PERF: use names/kinds for dtype inference on known types
Base   [87660ef] : Merge pull request #6968 from ahlmss/AHLMSS_0.13.1_ahl1

[hayd]

wow, quite a perf improvement!

[dsm054]

FWIW it worked for me to fix this symptom on my 64-bit box; my 32-bit isn't accessible, but I'll give it a go later, just for sanity purposes.

[jreback]

gr8; I merged into master. lmk if any other issues. thanks for the report!