BUG: fix time cast-safety for `factor*unit` e.g. in `10**6*ms == 1*s` (#19731)

* for `datetime64` units smaller than `s`, handle some
common/reasonable "metrix prefix" casting conversions like
1000 ns == 1 us

* this adds support specifically for equivalencies that span
10^3 and 10^6 fold, but there are even more beyond that, and if
we want to handle them all we may want to abstract to a handler
function

* that said, this still seems like a solid improvement that covers
some of the more common cases?

* MAINT: PR 19731 revisions

* add support/testing for datetime unit conversions at the
scale of 10^9, when below the 32-bit integer overflow limit

* raise a `TypeError` when `M8` unit magnitude exceeds
`INT_MAX`

* MAINT: PR 19731 revisions

* add new test `test_datetime_prefix_conversions()` that
performs bidirectional metric time prefix casting checks

* use `long long int` for `true_meta_num` on 32-bit systems

* some C code whitespace changes requested during review

* MAINT: PR 19731 revisions

* flake8-related fixes

* MAINT: PR 19731 revisions

* use a fixed size `int64_t` type for `true_meta_num`
on Windows to allow the overflow check to happen properly

* whitespace adjustments for the C code based on review comments

* MAINT: PR 19731 revisions

* improve the portability of the overflow
check for datetime unit integer prefixes

* MAINT: PR 19731 revisions

* simplify the changes by using `NumPyOS_strtoll` (less
custom C needed this way)

* MAINT: cleanup after merge conflict.

* MAINT: PR 19731 revisions

* based on reviewer feedback, make negative
datetime units a `TypeError` (including regression
test)

0 files changed, 0 insertions, 0 deletions