PERF: tz_convert/tz_convert_single

[jbrockmendel]

Making these follow the same pattern we use elsewhere, we get a perf bump:

In [2]: dti = pd.date_range("2016-01-01", periods=10000, tz="US/Pacific")        
In [3]: %timeit dti.tz_localize(None)                                                                                                                                                                              
91.4 µs ± 3.06 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)  # <-- PR
102 µs ± 6.43 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)  # <-- master

In [4]: ts = pd.Timestamp.now("US/Pacific")                                                                                                                                                                        
In [5]: %timeit ts.tz_localize(None)                                                                                                                                                                               
17.2 µs ± 307 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)   # <-- PR
19.6 µs ± 233 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)   # <-- PR

Next up is making sure we have full asv coverage for tz_convert/tz_convert_single, analogous to #35075. That can either be separate or added to this PR.

[simonjayhawkins]

is this test no longer relevant?

[jbrockmendel]

this PR requires that one of the two args be UTC

[simonjayhawkins]

hmm, that's what tz_convert_single was doing before? so what's the difference now between tz_convert and tz_convert_single? and is converting between 2 tzs no longer needed?

[jbrockmendel]

so what's the difference now between tz_convert and tz_convert_single?

One handles a scalar (i.e. Timestamp) the other handles an array. I'm trying to de-duplicate these further, havent found a way to do so without a perf hit.

and is converting between 2 tzs no longer needed?

correct.

[jreback]

looks fine, @simonjayhawkins comment