Spans block duplication

[nataxcan]

This PR implements block duplication.
Previously, span cache-hits were treated the same way prefix caching does it: same memory reference, different request being served.
But while the same memory reference contains the same semantic content (it refers to KV vectors of the same input tokens), it cannot contain two different positional encodings at once.
So, instead of using the same memory reference, we let vLLM allocate new blocks, but instead of prefilling those blocks we copy the KV vectors with adjusted (if needed) positional encodings.

[nataxcan]

currently running evals to check accuracy...

[starpit]

i see that this addresses the regression in SPANS_DEBUG. but the hashes printed out aren't useful. to turn bytes into a string, i think we need something like?

# just for example...
b = bytes.fromhex("abcd1234")

# turn b back into the string "abcd":
from binascii import hexlify
def pretty(b):
  return hexlify(b).decode('utf-8')[:4]

pretty(b)
abcd

whereas this PR uses str(b)[:4] which produces pretty much garbage output.