Add debugging code to gpu_dynamic_shared to try to track down hard-to-repro bug on buildbots

[steven-johnson]

No description provided.

[steven-johnson]

aha: upgrading the nvidia driver on my local linux box to a recent one (460.x) allows me to repro locally - the older driver I was using(440.x-ish) didn't fail. So it's driver-specific to some extent.

[steven-johnson]

If I run with cuda-memcheck, I get:

========= Invalid shared write of size 4
========= at 0x00000210 in _kernel_g_s0_x_x___block_id_x
========= by thread (0,0,0) in block (0,0,0)
========= Address 0x00001900 is out of bounds

[steven-johnson]

Bisect is suggesting that #5757 is the injection spot; investigating

[steven-johnson]

(1) The mystery hang seems to have resolved itself, or at least is no longer reproducing at recent top-of-tree.
(2) The bug with disable-llvm-opt enabled seems to be an unrelated red herring (this test fails with this flag enabled for as far back as I can find). See #5900 for a workaround there.