[WIP] Improve CAGRA-Q performance and add support for PQ_LEN=8

[enp1s0]

This PR:

1. Introduces the E5M2 data type as the internal storage type to improve performance
2. Adds support for PQ_LEN = 8 to achieve a higher compression ratio

E5M2 as smem data type

Using a lower-precision data type helps reduce shared memory bank conflicts and can improve throughput.
Since the quantization error from VQ+PQ is typically larger than the representation error of E5M2, the impact on search recall is expected to be negligible.

TODO: performance improvement figure

---

Support for PQ_LEN=8

The current cuVS implementation supports only PQ_LEN = 2 (4 bits per vector element) and 4 (2 bits per vector element).
This PR adds support for PQ_LEN = 8 to enable a higher compression ratio (1 bit per vector element).

TODO

• Reduce unnecessary compute_distance_vpq instances