Report write and read amplification

[venkat1701]

Closes #61

Stacked on #65, since both change the same harness files. Merge #65 first; this PR's base then moves to main once the #65 branch is deleted on merge.

The Resources table gets two new ratios per workload, plus bytes read:

Ratio	Physical bytes	Logical bytes (fixed estimate per workload)
write amplification	bytes written (from #38)	8 bytes per id plus the value: node-<i> for a new node, 24 bytes for an updated value, (cardinality + 1) × 8 for a new or deleted edge, 16 bytes per removed member
read amplification	bytes read from storage	8 bytes per id returned (the checksum counts them)

Correction (pushed after the PR was first opened): I first computed HStore's bytes read as pages read × page size, assuming each read takes a whole page. It doesn't: SegmentFile.read reads the first 4 KiB, then only the rest of the image if it's bigger. That overstated HStore's reads by up to 4 times. So this PR now also:

• adds PageStore.bytesRead and EngineStats.dataBytesRead, which count the bytes actually read
• has the adapter use that

JE reports sequential plus random read bytes. Both adapters carry the counts across reopen.

Report measures can now be computed from several fields, which the ratios need.

Smoke runs at scale 1, async, one run each, with the corrected count:

• With default caches HStore reads nothing from storage on warm reads, since its node cache holds everything.

• With CACHE_MB=64:

  	HStore	HyperGraphDB
  read.incidence per op	4 B	1.5 KiB
  read.incidence.cold per op	38 B	10.8 KiB
  read.members per op	4.4 KiB	84 B
  read.twohop per probe	34.9 KiB	1.8 KiB

  Member scans are the outlier: about 157 times the bytes they return, because every small edge has its own membership page. That's Small edges always get their own membership tree #55.

• Write amplification matches what bytes per operation already showed: HStore is 1.3 to 2× better on batched writes and large edges, and 6.6× worse on single commits (662 against 100), which is Small commits are slow and occasionally stall for seconds #46.

Tested:

• each commit compiles on its own; full ./mvnw install passes
• new EngineTest.bytesReadCountWhatComesOffTheSegmentFiles: after reopen, bytes read are at least a header and at most a page per page read
• old result files render unchanged
• results without the new fields render without the new rows