f16/half-float support

[gnzlbg]

I was implementing the v{min,max}v ARM and AArch64 intrinsics and ran into a couple of them that require passing <4 x half> vectors to llvm and returning half types:

• pub unsafe fn vmaxv_f16 (a: f16x4_t a) -> f16
• pub unsafe fn vmaxvq_f16 (a: f6x8_t a) -> f16
• pub unsafe fn vminv_f16 (a: f16x4_t a) -> f16
• pub unsafe fn vminvq_f16 (a: f16x8_t a) -> f16

For example, vmaxv_f16 should map to: call half @llvm.aarch64.neon.fminv.f16.v4f16(<4 x half>)

[alexcrichton]

FWIW I think we've got basically no trans/language support for f16 right now, so I'd just leave them out

[gnzlbg]

The PowerPC 64 architecture also supports a vector _Float16 type, with 16 16-bit floats.

[tgross35]

f16 and f128 are now unstably available, so these can be added. This extends to platforms beyond Arm, see rust-lang/rust#125440 for my primitive list