perf: snapshots: use TextEncoder instead of new Blob to fix heavy GC lag

[llc0930]

Tested for over a month in an environment with 9,000+ tabs and the Firefox profile on a traditional HDD, with a noticeable reduction in perceived lag when deleting snapshots.

[mbnuqw]

I profiled both methods (Blob and TextEncoder) with a non-ASCII character set (Zh): Both methods produced approx the same memory footprints. Both methods triggered string re-encoding (x2 CPU usage compared to ASCII-only chars) + memcpy. And they showed more or less the same GC behavior. The Blob method has a bit lower memory threshold that triggers GC, which probably causes this effect. But if the data size continue to grow, you'll start experiencing heavy GC lag with TextEncoder too.

To fix this we can pre-allocate and keep some memory for measuring (e.g. 10mb). But it brings two trade-offs:

• size measurement will be strictly limited to this 10mb, which is ok, there is a hard limit of 10mb on all snapshots anyway (or we could dynamically re-create this buffer, but this will trigger GC...).
• this 10mb of RAM will always be consumed by Sidebery. I think we can afford such luxury even these days.

btw, this size measurement will be removed when I migrate snapshots storage to OPFS.