revert(plugin): drop the per-call plugin deadline

[JeroenSoeters]

Summary

The retry-aware watchdog work bounded every watched plugin call by a deadline the agent supplied through the operator's spawn environment. This removes it and keeps the rest of that work.

The deadline never fixed the reported failure. Healthy resources were declared missing in action because one plugin RPC blocked inside its own retry loop far longer than the status-poll interval, starving the loop that produces heartbeats. Bounding that retry in the AWS plugin is what fixed it. The agent-side deadline only enforced the invariant the plugin had violated. Both workflow regression tests for the reported bug pass without it.

What it cost was disproportionate. A deadline is not an additive interface change a plugin author can ignore, the way an added request field is. It is a cross-cutting obligation that applies inside every method, is violated by omission rather than commission, fails silently, and cannot be verified by reading your own code. It also cannot be enforced: Go has no goroutine preemption, so a plugin that does not thread the context still blocks and the agent still waits. Paying that much interface surface for a guarantee that is not one is the wrong trade, particularly as plugins are increasingly written by agents that implement an interface faithfully and observe a prose invariant unreliably.

Kept

• The watchdog window derived from the operator's actual retry cadence, including the flat RetryDelay term.
• Status calls no longer terminating without sending a final progress, so the real error reaches the updater instead of a watchdog message a minute later.
• The window sized from the per-plugin config the operator was really spawned with, rather than the deprecated global.

Changed

The window keeps a per-call allowance, renamed PluginCallAllowance and documented for what it actually is: the silence the agent tolerates for one call, matching the longest the updater itself waits for a reply, rather than a bound anything enforces. The agent's own outer call timeout returns to 60s, since there is no longer an operator deadline for it to outlast.

TestFailedStatusCallNeverReissuesTheOriginalOperation is kept and narrowed to the recoverable path, which is the only one that can reschedule. The plugin double it shared with the deleted deadline tests moves to plugin_operator_double_test.go.

Verification

go build ./..., the full -tags=unit suite across internal/... and pkg/plugin, and the three MIA workflow tests under -tags=integration (TestSlowHeartbeatIsNotDeclaredMissingInAction, TestFailedStatusCheckReportsThePluginError, TestPluginOperatorCrashConvergesViaTimeout) all pass. That the first two still pass without the deadline is the evidence that the window sizing, not the deadline, is what fixed the reported bug.