Reduce use of `span.with_lo(span.lo() + BytePos(1))` in `librustc_parse/parser/mod.rs`

[JohnTitor]

span.with_lo(span.lo() + BytePos(1)) may cause an ICE if span points strange character so we should use next_point instead. But next_point isn't actually equivalent with span.with_lo(span.lo() + BytePos(1)) so it breaks some spans in current diagnostics. We should find out the cause and reduce use of them.

See context: #68735 (comment)

[dwrensha]

I think #68845 should fix this.