Fix Hilbert filter: quadrature branch used the in-phase (I) taps

[spinkham]

Summary

Hilbert::filter_block (Delay.cpp) builds both the real and imaginary parts
of the analytic output from the in-phase coefficient table (IHilbertBPFirCoef).
The quadrature pointer QKptr is initialised but never used: the queue stores the
real input duplicated into both cmplx lanes, and the inner loop multiplies the
whole cmplx sample by the scalar I coefficient, so I and Q come out identical —
a 45°-rotated real signal rather than a 90° Hilbert pair.

Because the "analytic" input to the fading stage is not truly analytic, the
Rayleigh fading it produces is not circular. For a CW probe faded through a 2-path
CCIR-Poor channel the output envelope is Hoyt/Nakagami-like (mean/rms ≈ 0.78,
roughly double the deep-fade probability of Rayleigh) instead of the Rayleigh
value 0.886.

Fix

Apply the I taps to the real accumulation and the Q taps (the already-computed
QKptr) to the imaginary accumulation:

cmplx acc{0., 0.};
for (int j = 0; j < HILBPFIR_LENGTH; ++ j, ++ Firptr) {
    acc.r += Firptr->r * (*IKptr++);   // in-phase: band-pass (I) taps
    acc.i += Firptr->i * (*QKptr++);   // quadrature: Hilbert (Q) taps
}
pOut[i] = acc;

Verification

1500 Hz CW probe faded through --spread 1.0 --delay2 2 --spread2 1.0:

metric	before	after	Rayleigh ref
analytic input |z| std/mean (0 = ideal)	0.52	0.02	0
faded envelope mean/rms	0.784	0.896	0.886

After the fix the faded-envelope distribution matches Rayleigh across the band.

Note: this was verified on a byte-identical Delay.cpp (the Rhizomatica tree),
not a fresh build of this repo — master currently fails to build on GCC-13 due
to a missing <cstring> include, which is unrelated to this change.