Fix issues seen on devcloud with flex170 gpu

[fcharras]

• pin numba_dpex build from dppy/label/dev conda channel
• use a default hardcoded value when device global memory cache size can't be fetched

For the latter fix it will be interesting to test during the next session on the devcloud if clinfo also fails at finding global_memory_cache_size, else we could use pyopencl.

[fcharras]

TY @jjerphan I've applied your suggestions

I also just noticed while updating the TopK code, that I put wrong numbers in the predicted local memory usage of the kmeans kernels (kmeans++, lloyd and compute_labels), which is needed to run on CPU.

When fixing it I also saw that sometimes SYCL refuses the kernel call if the predicted local memory is too close from the maximum amount of available local memory, so I also ensured to leave a free buffer. (probably that the device itself use it to optimize and load things)

It's a small diff so I took the liberty to hijack it into the last commit.

[fcharras]

TY for review @jjerphan