Keep the sign in integer powers of negative numbers

[timesselens]

On the device, @fastmath x^n with x::Float32 or Float16 and an integer n outside -1:3 returned NaN for negative x. Literal exponents were affected too, since @fastmath x^7 goes through pow_fast(x, Val(7)) and then the same method.

k!(y, x, n) = (i = threadIdx().x; @inbounds y[i] = @fastmath x[i]^n; nothing)
x = CuArray(Float32[1.5, -0.5, 3, -2]); y = similar(x)
@cuda threads=4 k!(y, x, 7); Array(y)
# main:    Float32[17.085943, NaN, 2186.9993, NaN]
# this PR: Float32[17.085943, -0.007812505, 2186.9993, -128.0]

The override from #3098 calls __nv_fast_powf(x, Float32(n)), which computes exp2(n * log2(x)). This PR computes the power of abs(x) and copies the sign of x when n is odd, the same way LLVM's AMDGPU back-end expands a fast pown (AMDGPULibCalls.cpp). The parity comes from the integer n, because Float32(n) can round an odd exponent to an even one (16_777_217). With a literal exponent the sign handling compiles to a single copysign.f32, and positive bases give the same results as before.

The precise ^(::Float32/Float64, ::Integer) from #3218 had the same rounding problem: (-1f0)^16_777_217 returned 1. The second commit takes the sign from the integer exponent there as well.

Alternatives considered: __nv_powif (what CUDA C's pow(float, int) uses, and the equivalent of the CPU's llvm.powi) matches the CPU @fastmath result bit for bit and is more accurate (for example 0 against 5 ulp on (-0.5f0)^7, and 153 against 1726 ulp on 1.001f0^5000). But with an exponent that is only known at run time it is 1.6× slower for n = 7 and 5× slower for n = 1000 (RTX 5080), so pow_fast keeps the fast path.

[maleadt]

LGTM, thanks! I also applied this to the non-fastmath ^(::Float32/Float64, ::Integer). And some other minor improvements: switched to copysign, which compiles to a single instruction for a literal exponent, and made the tests compare element by element, since the array-norm ≈ hid a Float16 sign flip.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.90%. Comparing base (f8606eb) to head (d284892).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3301   +/-   ##
=======================================
  Coverage   85.89%   85.90%           
=======================================
  Files         187      187           
  Lines       19022    19025    +3     
=======================================
+ Hits        16339    16343    +4     
+ Misses       2683     2682    -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.