Fix the VMX backend

[pkubaj]

The ICBC_VMX path had evidently never been compiled: "vectro float", a
vec_ld() call with the wrong arity and no terminating semicolon, vec_div()
against an integer splat (there is no integer vector divide on PowerPC, so
this is ambiguous rather than promoted), and a compound literal that VFloat
has no constructor for. all(), any() and mask() were missing from the
backend entirely; mask() is used by compute_indices3() and
compute_indices4().

vload() now uses vec_xl() where VSX is available. vec_ld() is lvx, which
masks off the low four bits of the address and therefore quietly returns
the containing aligned vector for an unaligned pointer.

mask() follows the NEON implementation, since PowerPC has no movemask:
test each lane's sign bit, AND with {1,2,4,8}, then OR the lanes together.

Verified on FreeBSD/powerpc64le with clang 19: compress_bc1() over 20000
pseudo-random blocks gives byte-identical output and an identical total
error for ICBC_SIMD=ICBC_VMX and ICBC_SIMD=ICBC_SCALAR.