fix implementation of NOR polling APIs

[drossetti]

This was observed during a code review and needs to be thoroughly tested.
Note that NOR support is currently disabled by default and is only supported on Volta+.

[drossetti]

@e-ago ping for this PR

[e-ago]

Considering:

• a modified version of libmlx5 (according to IBV_EXP_PEER_OP_POLL_NOR_DWORD_CAP not working correctly libmlx5#7)
• the gds_kernel_latency test with !peersync (-P -U options)
• the NOR operation is IBV_EXP_PEER_OP_POLL_NOR_DWORD

With this PR the test hangs and returns several errors:

...
ERR:  [main] [63274] unexpected tx ev 0, batch len 20
ERR:  [main] [63274] unexpected rx ev 1, batch len 20
ERR:  [main] [63375] unexpected tx ev 0, batch len 20
ERR:  [main] [63375] unexpected rx ev 1, batch len 20
ERR:  [main] [63453] unexpected tx ev 0, batch len 20
ERR:  [main] [63453] unexpected rx ev 1, batch len 20
ERR:  [main] [63543] unexpected tx ev 0, batch len 20
ERR:  [main] [63543] unexpected rx ev 1, batch len 20
ERR:  [main] [48] unexpected rx ev 0, batch len 20
ERR:  [main] [49] unexpected rx ev 0, batch len 20
ERR:  [main] [50] unexpected rx ev 0, batch len 20
...

Without this PR, everything works correctly.
You should change done = (0 == ~(data | value)); into done = (0 != ~(data | value));

[drossetti]

@e-ago you are right, I misinterpreted the HW spec once again. So here the patch simply makes the condition more explicit. Sorry for the noise.