SBT record size is overly conservative

[natevm]

At the moment, we just use 4096 bytes per SBT record. But if we have 1 million geometries, that's 4GB of memory... Since many of these records are actually not so big, especially geometry records. So we might A) consider if we can have different record sizes for hit groups vs miss vs raygen, but then also B) how to efficiently determine what the smallest acceptable record size is.