Fix Metaspace decoder dropping internal spaces in the first token

[sohumt123]

What does this PR do?

Fixes the Metaspace decoder silently dropping internal word-boundary spaces in the first decoded token.

Bug

Metaspace::decode_chain is supposed to strip the single leading replacement char (▁) that pre-tokenization prepended (when prepend_scheme != Never) and turn every other replacement char into a space. Instead, for token i == 0 the closure mapped every occurrence of the replacement char to None:

if c == self.replacement {
    if i == 0 && self.prepend_scheme != PrependScheme::Never {
        None          // deletes ALL ▁ in the first token, not just the prepended one
    } else {
        Some(' ')
    }
}

So any first token containing more than one metaspace char loses its internal spaces:

• default split=true, with a BPE-merged token spanning a space: decode_chain(["▁in▁the", "▁house"]) → "inthe house" instead of "in the house"
• split=false on multi-word input: decode_chain(["▁Hey▁friend"]) → "Heyfriend" instead of "Hey friend"

Fix

Track the char position within the token and drop the replacement char only at position 0 of the first token; every other occurrence decodes to a space. This matches the SentencePiece convention (replace all ▁ with spaces, strip exactly the one prepended prefix space) and the transformers slow-tokenizer behavior.

How tested

• Added decode_first_token_with_internal_replacement covering split=true, split=false, and PrependScheme::Never (unchanged: nothing stripped). The test fails on main (left: ["inthe", " house"], right: ["in the", " house"]) and passes with the fix.
• Full rust suite: cargo test --manifest-path ./tokenizers/Cargo.toml --lib → 202 passed, 0 failed. cargo clippy --all-targets --all-features -- -D warnings and cargo fmt --check clean.
• Existing decode tests for Always/Never schemes (single leading ▁ first tokens) are unaffected, as are the Python-binding Metaspace decode expectations.