telemetry: count and correctly report samples dropped when account is full

[elitegreg]

Resolves: #4127

Stacked on #4144 (gm/telemetry-drop-samples-metric), which adds the shared samples_dropped_total counter this reuses. Base is that branch, so the diff shown is just this change; merge #4144 first.

Summary of Changes

• Account-full drops are now visible in metrics. SubmitSamples returns nil on ErrSamplesAccountFull so Tick counts it as a successful submission, which means data loss previously left nothing but a warning. It now increments doublezero_device_telemetry_agent_samples_dropped_total{reason="account_full"} and submitter_account_full on the errors counter.
• The logged count is fixed. The old warning reported len(samples), the whole flushed partition, rather than what was actually lost, overstating the loss whenever a partition spans more than one batch (a 300-sample partition that filled the account on its second batch logged droppedSamples=300 when 61 samples were lost).
• The two duplicated account-full blocks (direct write and post-initialize) are factored into handleAccountFull.
• Control flow is unchanged: the account cannot accept further writes, so returning an error would only buy five pointless retries. The partition is still removed and the caller still treats it as done.

Note on the dropped count

The issue proposed counting len(batch). That undercounts for the same reason len(samples) overcounts: when the account fills mid-partition, the batches behind the failing one are never attempted and are discarded along with it, and Buffer.Remove also throws away whatever the collector buffered since the flush. handleAccountFull therefore counts len(samples)-i (the failing batch plus everything after it) and adds the partition's buffered length, read before the removal. The log reports the two separately as unsubmittedSamples and bufferedSamples.

Diff Breakdown

Category	Files	Lines (+/-)	Net
Tests	1	+140 / -0	+140
Core logic	1	+23 / -4	+19
Scaffolding	1	+3 / -1	+2
Docs	1	+1 / -0	+1
Total	4	+167 / -5	+162

Mostly tests; the fix itself is one extracted helper and two call sites.

Key files (click to expand)

• controlplane/telemetry/internal/telemetry/submitter.go — new handleAccountFull records the error and drop counters, logs the unsubmitted and buffered counts, and removes the partition; both ErrSamplesAccountFull branches in SubmitSamples now call it with len(samples)-i
• controlplane/telemetry/internal/metrics/metrics.go — adds the account_full drop reason and the submitter_account_full error type

Testing Verification

• New test adds a sample from inside the write callback (after the flush, so it is only reachable via Buffer.Remove) and asserts the drop counter rises by 3 for a 2-sample flush, the error counter by 1, and the log carries unsubmittedSamples=2 bufferedSamples=1.
• New multi-batch test flushes 300 samples with the first batch succeeding and the second hitting a full account, and asserts exactly 61 drops and no unsubmittedSamples=300. Against the pre-fix submitter this test sees the old droppedSamples=300, confirming the miscount.
• New test covers the post-initialize account-full branch, which the extracted helper now shares.
• All three fail against the pre-fix submitter. Existing account-full tests (drops_samples_if_account_full, initializes_then_drops_samples_if_account_full) still pass unchanged.
• The new subtests run serially and compare counter deltas, since the metrics are package-level and shared with those existing parallel subtests.