perf(lsp): use string.buffer for rpc loop

Avoids some table allocations.
In a quick test over 50000 iterations it reduces the time from 130ms to
74 ms

For the test setup details see:

https://github.com/mfussenegger/nvim-dap/pull/1394#issue-2725352391

runtime/lua/vim/lsp/rpc.lua

@@ -25,8 +25,8 @@ local function get_content_length(header)
     if line == '' then
       break
     end
-    local key, value = line:match('^%s*(%S+)%s*:%s*(.+)%s*$')
+    local key, value = line:match('^%s*(%S+)%s*:%s*(%d+)%s*$')
-    if key:lower() == 'content-length' then
+    if key and key:lower() == 'content-length' then
       return tonumber(value)
     end
   end
@@ -39,8 +39,37 @@ local header_start_pattern = ('content'):gsub('%w', function(c)
   return '[' .. c .. c:upper() .. ']'
 end)
 
+local has_strbuffer, strbuffer = pcall(require, 'string.buffer')
+
 --- The actual workhorse.
-local function request_parser_loop()
+---@type function
+local request_parser_loop
+
+if has_strbuffer then
+  request_parser_loop = function()
+    local buf = strbuffer.new()
+    while true do
+      local msg = buf:tostring()
+      local header_end = msg:find('\r\n\r\n', 1, true)
+      if header_end then
+        local header = buf:get(header_end + 1)
+        buf:skip(2) -- skip past header boundary
+        local content_length = get_content_length(header)
+        while #buf < content_length do
+          local chunk = coroutine.yield()
+          buf:put(chunk)
+        end
+        local body = buf:get(content_length)
+        local chunk = coroutine.yield(body)
+        buf:put(chunk)
+      else
+        local chunk = coroutine.yield()
+        buf:put(chunk)
+      end
+    end
+  end
+else
+  request_parser_loop = function()
     local buffer = '' -- only for header part
     while true do
       -- A message can only be complete if it has a double CRLF and also the full
@@ -102,6 +131,7 @@ local function request_parser_loop()
       end
     end
   end
+end
 
 local M = {}