perf: intern per-character Nodes in characters()

[AmitMY]

The characters() unit function allocated a fresh Node for every character. This mirrors the utf8 byte layer before #36 — share one immutable Node per character instead, via a lazy cache (the unbounded analogue of _BYTE_NODES):

@lru_cache(maxsize=None)
def _char_node(char: str) -> Node:
    return Node(char.encode("utf-8"))

Nodes are frozen, so sharing is safe, and unlike caching NodesSequence/graph builders there's no memo/GraphSettings dependency (a Node is just bytes). Wins the same two ways #36 did:

• Memory — repeated characters collapse to shared objects.
• Speed — equal characters are the same object, so merge's nodes[i:i+m] == merge comparisons hit CPython's per-element identity short-circuit.

Output identical, 137 tests pass. Measured on repeated multilingual text with units="characters" (best of 3):

	main	this PR
time	0.114s	0.073s (−36%)
peak mem	2.26MB	1.25MB (−45%)

Only affects the characters unit path (not the default utf8_clusters), so the default BPE benchmark is unchanged — but for codepoint-level tokenization it's a sizable win.

🤖 Generated with Claude Code