[alpaka] change assert with ALPAKA_ASSERT_OFFLOAD

[antoniopetre]

Some Alpaka kernels have the occupancy lower than the Native CUDA implementation (for example countFromVector kernel from HistoContainer).

I found out that the problem was related to the assert use in Alpaka. In Native CUDA, the assert is disabled by default, becoming enabled only if the macro GPU_DEBUG is set (this is not the case in Alpaka). I changed the assert from the Alpaka implementation with the macro ALPAKA_ASSERT_OFFLOAD from the Alpaka library (compatible from Alpaka 0.6.1). This macro enables assert only if ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST is set. This change increased the occupancy for some kernels in Alpaka.

[makortel]

Given that cudacompat has assert() enabled, should we include ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST by default for serial and TBB backends?

@fwyzard

[makortel]

I'm going to merge this as it is, and I opened an issue #228 to remind of the question in #219 (comment).