feat: supply chain pack supplier domain

[denironyx]

Generated register-shaped entities and export them for entity resolution Addresses, company names and people are now generated from distributions measured on two real company registers rather than from Faker defaults:
addresses are shared on a Zipf curve, names carry the register's legal forms and spellings, and 4.5% of companies have anybody attached at all.

graphfaker register reads a register in Senzing format from a file, directory or zip, and --sink senzing writes a generated one back out with senzing/entities.parquet as the answer key. Scored with graphfaker.resolve: the name alone gives 0.93 F1, adding the address costs recall, and matching on address alone gives 0.03 precision.

Corrects an earlier claim: the 86.7% name collision rate counted records rather than companies, so it conflated one company written many times with two companies sharing a name. Per company it is 1%.

Also fixes: gen --export crashed on a Windows console, the DuckDB property graph failed on a dataset with no edges, and the benchmark's name normaliser never stripped L.L.C.