LSP is using Line & column as symbol position, clangd needs to transfer file
offset to Line & column when sending results back to LSP client, which is a high
cost, especially for finding workspace symbol -- we have to read the file
content from disk (if it isn't loaded in memory).
Saving these information in the index will make the clangd life eaiser.
There are 4 of these per symbol, if we can keep line + character to 32 bits we save 16 bytes per symbol.
That looks like it might still be ~10% of non-string size. WDYT?
(12 bits for column and 20 bits for line seems like a reasonable guess)