ep_perfsuite: latency/bandwidth tests use wrong DCP sequence numbers

The ep_perfsuite tests implemented using perf_dcp_latency_and_bandwidth are currently broken as they do not correctly handle DCP sequence numbers. The affected tests are:

• DCP latency (Padded JSON)
• DCP latency (Random JSON)
• DCP latency (Random BIN)

These tests all run in two stages:

• First they load load documents "as-is" (compression off), and create a DCP stream without compression and measure latencies / sizes.
• They then run a second time, loading the same documents again but this time with a DCP stream with compression enabled.

The issue is that these all use the same bucket, and the DCP streams do /not/ take into account sequence numbers - therefore the second DCP stream receives 2x the items (all the "as-is" documents in addition to it's own.

The test needs to be fixed - the simplest approach is to probably just make the two phases two totally separate tests (i.e. start with an empty bucket each time).

Note this issue essentially blocks MB-18670 - to fix that I need to refactor how the end of stream is found, and this uncovered the fact that these tests are currently broken - and having them "re-play" the first set of mutations means the sentinel approach doesn't work.