Add Token.as_str() to decode bytes token values

[apoorva-01]

float(token) breaks with use_bytes=True: could not convert string to float: Token('NUMBER', b'1'). Token subclasses str, but under use_bytes the value is bytes, so the string is the repr "b'1'", not "1".

Added Token.as_str() that decodes bytes (else returns as-is), so float(token.as_str()) works either way. Used latin-1 since that's what lark uses for lexer patterns and it won't choke on odd bytes. Can switch if you want.

This is the small fix. It doesn't touch the deeper thing where token.value == token is false under use_bytes. Making Token bytes-aware would fix that too, but it's a bigger semantics change, so I kept it contained. Happy to go that way instead. Test parses bytes and checks float(token.as_str()), fails without the fix.

Closes #1526