t27b: std.math.pow(f32, ..) from libm.t27 (Zig's std/math/pow.zig); gf16.t27 passes #7823

Part of #7746 (t27b passes specs/numeric/gf16.t27), which is part of #6063 (t27b coverage), item C10 of #6488. This is the last PR of #7746's split, after #7791, #7819 and #7822. It makes gf16.t27 pass and closes #7746.

What gf16.t27 needs

gf16_pow calls std.math.pow(f32, base_val, exp_val), and gf16_pow_two_cubed and others reach it. After #7822 it is gf16.t27's last blocker (ExprCall(std.*), call to std.math.pow).

What the reference does

std.math.pow(f32, x, y) is not compiler_rt. It is Zig 0.16's std/math/pow.zig, a port of Go's math.Pow:

• The special cases come first: y = +-0, x = 1, a NaN, y = 1, x = +-0, y = +-inf and x = +-inf.
• y == 0.5 is @sqrt(x).
• Otherwise y is split by math.modf (an f32 @trunc). The fractional part gives @exp(yf * @log(x)), the f32 expf and logf. The integer part is a square-and-multiply on math.frexp(x), and the result is math.scalbn(a1, ae), which is math.ldexp, with its own rounding of a subnormal result.
• isOddInteger takes an f64 and calls the f64 math.modf.
• A NaN result of the special cases is math.nan(f32), 0x7FC00000.
• Measured on the t27c lab: a comptime-folded std.math.pow(f32, ..) equals the run-time call on a 49-point grid.

Plan

• Port pow for an f32 to t27 in specs/tri/t27b/libm.t27, with what it calls: frexp and ldexp of an f32, isOddInteger, and compiler_rt's f64 trunc. It reuses the ports of expf, logf (t27b: f32 @log, @log2 and @log10 from libm.t27 (compiler_rt logf, log2f, log10f) #7822), truncf (t27b: f32 @floor, @ceil, @round, @trunc and @rem from libm.t27 (compiler_rt floorf, ceilf, roundf, truncf, fmodf) #7819) and nan.
• specs/tri/t27b/libm_plan.t27: std.math.pow names its float type and takes three operands. An f32 calls powf, and an f64 is refused.
• Hold the port to std.math.pow bit for bit over grids of x and y, in a spec of its own.

Acceptance criteria

• t27b test --blockers specs/numeric/gf16.t27 reports nothing.
• t27b test specs/numeric/gf16.t27 passes every test that t27c test-report passes, with the same verdicts.
• The lab corpus receipt compare gives IMPROVED_ONLY, EQUIVALENT or NEUTRAL.

User Scenarios

• Given the specs/numeric/gf16.t27 spec calls std.math.pow(f32, ...) via gf16_pow, When t27b test --blockers specs/numeric/gf16.t27 is run, Then it reports no blockers for std.math.pow.
• Given the specs/tri/t27b/libm.t27 file implements powf for f32 with frexp, ldexp, isOddInteger, and trunc from compiler_rt, When t27b test specs/numeric/gf16.t27 runs, Then every test passes with the same verdicts as t27c test-report.

Requirements

• FR-001: specs/tri/t27b/libm.t27 MUST implement powf for f32 matching Zig 0.16's std/math/pow.zig special cases and algorithm.
• FR-002: specs/tri/t27b/libm.t27 MUST provide frexp and ldexp for f32, isOddInteger using f64 modf, and use compiler_rt's f64 trunc.
• FR-003: specs/tri/t27b/libm_plan.t27 MUST declare std.math.pow as taking three operands (type, base, exponent), route f32 to powf, and refuse f64.
• FR-004: The implementation MUST produce bit-for-bit identical results to the reference over grids of x and y values.

Boundary

• specs/numeric/gf16.t27
• specs/tri/t27b/libm.t27
• specs/tri/t27b/libm_plan.t27