slog: faster and less memory-consumption (#28621)

These changes improves the performance of the non-coloured terminal formatting, _quite a lot_. 

```
name               old time/op    new time/op    delta
TerminalHandler-8    10.2µs ±15%     5.4µs ± 9%  -47.02%  (p=0.008 n=5+5)

name               old alloc/op   new alloc/op   delta
TerminalHandler-8    2.17kB ± 0%    0.40kB ± 0%  -81.46%  (p=0.008 n=5+5)

name               old allocs/op  new allocs/op  delta
TerminalHandler-8      33.0 ± 0%       5.0 ± 0%  -84.85%  (p=0.008 n=5+5)
```

I tried to _somewhat_ organize the commits, but the it might still be a bit chaotic. Some core insights: 

- The function `terminalHandler.Handl` uses a mutex, and writes all output immediately to 'upstream'. Thus, it can reuse a scratch-buffer every time. 
- This buffer can be propagated internally, making all the internal formatters either write directly to it,
- OR, make  use of the `tmp := buf.AvailableBuffer()` in some cases, where a byte buffer "extra capacity" can be temporarily used. 
- The `slog` package  uses `Attr` by value. It makes sense to minimize operating on them, since iterating / collecting into a new slice, iterating again etc causes copy-on-heap. Better to operate on them only once. 
- If we want to do padding, it's better to copy from a constant `space`-buffer than to invoke `bytes.Repeat` every single time.

internal/testlog/testlog.go

@@ -183,8 +183,7 @@ func (h *bufHandler) terminalFormat(r slog.Record) string {
 	}
 
 	for _, attr := range attrs {
-		rawVal := attr.Value.Any()
-		fmt.Fprintf(buf, " %s=%s", attr.Key, log.FormatLogfmtValue(rawVal, true))
+		fmt.Fprintf(buf, " %s=%s", attr.Key, string(log.FormatSlogValue(attr.Value, true, nil)))
 	}
 	buf.WriteByte('\n')
 	return buf.String()