IndexIVFRaBitQFastScan: add high-accuracy LUT/accumulation path for multibit RaBitQ+ search

[lyang24]

Summary

a higher-accuracy LUT / accumulation path so the extra bits can be used effectively during FastScan search.

Proposal

Add a high-accuracy FastScan path for IndexIVFRaBitQFastScan that can be used when multibit RaBitQ+ is enabled.

Possible design points:

• keep the current baseline path for the 1-bit or low-precision case
• add a separate higher-accuracy LUT transformation path
• use a higher-accuracy accumulation kernel where required
• select the path explicitly or automatically based on nb_bits

[mnorris11]

cc @alibeklfc