Float controls, then narrow wavefronts

Two of the three pieces asked for. 592 workspace tests green.

The division sequence is complete (D144). The pre-scale needed no float controls in the end - and could not have had them, because this device does not support preserving 32-bit subnormals at all.

Thirty-two-lane shaders translate (D145), with the width carried on the strategy.

Surprises

• The device answered the question by saying no. The first thing the new device report printed was subnormals flushed. The plan had been to request preservation; the plan was impossible. It turned out not to be needed: zero and subnormal share an all-zero exponent field, and everywhere the question is asked the true value cannot be zero - so one test on the bits is correct on a flushing device and a preserving one alike.
• The report also settled the next task's main parameter. Host subgroup width is 32 and the guest wavefront is 64, so a subgroup mapping is 2:1 rather than 1:1 - which is exactly what made narrow-wavefront support worth doing before it.
• A narrow shader is not the same shader with fewer lanes. It uses different instructions for its masks, and those were landing in the register file instead of the mask. Every lane would have stayed active for the whole shader.
• My own test asserted the wrong thing about lanes. The pre-scale's flag came back as all ones rather than 1, because every lane gets the same operands and every lane flags. Asserting 1 would have been asserting the other sixty-three lanes did not run.