REL: Prepare for NumPy 1.20.0rc2 release.
MAINT, BUG: Adapt `castingimpl.casting` to denote a minimal level
NEP: Accept NEP 35 as final
This also allows skipping the call to `resolve_descriptors()` when the casting level guarantees that the cast will be successfull. This is sufficient to "tape over" problems with object to string casts, because it means that an "unsafe" cast is known to work (no matter the string length). Note that there is a theoretical anomaly remaining: Casting object to string without first discovering the correct string length will fail (because it is unclear which string length to pick), even though no matter which string length is picked "unsafe" is correct. NumPy should never do this, we always first discover the string length when casting from `object` DType specifically (object is special that way -- not string itself). In general, this seems fine to me, in practice, the behaviour would probably not be quite fixed yet.
This accepts NEP 35 as final. There has been no discussion about it in a long time. The current mode is strict about type input (`like=` must be an array-like). So that most of the "open" points are OK to remain open. Unless we need to discuss the name `like` or the fact that we pass an array-like itself, the previously noted open points gh-17075 all seem not very relevant anymore.
BUG: Add -std=c99 to intel icc compiler flags on linux
BUG, SIMD: Fix detect host/native CPU features on ICC at compile-time
Numpy will not build if this is not set because it has code that follows the c99 standard. icc is set to c89 by default. Look below at IntelCCompilerW which is the icc equivalent on Windows and that one already has c99 set.
Unlike GCC and CLANG, Intel Compiler exposes all supported intrinsics, whether or not the build options for those features are specified. Therefore, we must test #definitions of CPU features when option native/host is enabled via `--cpu-baseline` or through env var `CFLAGS` otherwise the test will be broken and leads to enable all possible features.