[ntuple] Heuristically reduce memory usage of buffered writing

[hahnjo]

At the moment, if implicit multi-threading is turned on, RPageSinkBuf::CommitPage will unconditionally copy the (uncompressed) page and create a task for sealing:

root/tree/ntuple/v7/src/RPageSinkBuf.cxx

Lines 194 to 203 in a2cb0a6

	// TODO avoid frequent (de)allocations by holding on to allocated buffers in RColumnBuf
	zipItem.fPage = fPageAllocator->NewPage(page.GetElementSize(), page.GetNElements());
	// make sure the page is aware of how many elements it will have
	zipItem.fPage.GrowUnchecked(page.GetNElements());
	memcpy(zipItem.fPage.GetBuffer(), page.GetBuffer(), page.GetNBytes());
	fCounters->fParallelZip.SetValue(1);
	// Thread safety: Each thread works on a distinct zipItem which owns its
	// compression buffer.
	fTaskScheduler->AddTask([this, &zipItem, &sealedPage, &element, allocateBuf, shrinkSealedPage] {

This leads to increased memory usage for high compression ratios when no spare threads are available or too slow to keep up consuming the tasks. Heuristically, we could keep track of the total number of "queued" bytes and seal a page immediately if it is higher than a threshold (approx. zipped cluster size?). This should always leave enough work for other threads to pick up, while reducing memory usage because only the compressed page is kept.