fix(fill): re-anchor start_block under --no-reset-between-tests

[skylenet]

Problem

--no-reset-between-tests keeps the chain moving between fills, but the
session anchor never moves with it.

_session_pre_run captures start_block once, from latest, after the
global setup:

@pytest.fixture(scope="session", autouse=True)
def _session_pre_run(...):
    ...
    # 4. Capture start block (head after global setup).
    start_block = eth_rpc.get_block_by_number("latest")
    client_backend.start_block = start_block

That is the only production write to client_backend.start_block. Every
test then chains its first block from it, in make_stateful_fixture.

Normally the invariant holds because _reset_chain_between_tests rewinds
the chain back to that block after each test. Under
--no-reset-between-tests the rewind is skipped and nothing takes its
place, so the anchor goes stale as soon as the first test builds a block.

The nonce side of the flag was already adapted — worker_key reads at
latest under the flag, with a comment explaining the inversion. The
anchor side was missed.

Symptoms

Two failures with the same root cause. Which one appears depends on where
the client's state-history window sits:

where it fails	error
worker_key -> get_account	historical state ... is not available
build_block (first block of a later test)	parentHash is not current head

Seen while filling
tests/benchmark/stateful/bloatnet/test_setup_contracts.py on a
jochemnet snapshot. Both code_size parametrizations run in one session;
the first passes and advances the chain about 7,700 blocks, and the
second dies on the anchor it never updated.

Fix

Read the head before each test and re-anchor on it, so every test starts
from what the previous test actually left behind. That is the chain the
accumulating mode is built around.

Consumer impact

None. Consumers route pre-run setup by directory (pre_run/*.json,
applied once per session), not by matching a fixture's
start_block_hash. Each test's fixture now records the anchor it truly
chains from, which is more accurate than every test claiming the session
anchor while depending on its predecessors.

Testing

• ruff check and ruff format --check pass in repo context.
• pytest --collect-only over tests/benchmark/stateful/bloatnet
  collects 382 tests, unchanged from the base commit.
• Behaviour verified against the failing fill described above.