fix: correct softirq_latency metric behavior

softirq_raise can be triggered by hard interrupts and is re-entrant.
The previous implementation updated the timestamp on every raise, causing the reported value to reflect the minimum latency, which is misleading and of limited usefulness.

This change introduces a per-softirq guard flag. The timestamp is recorded only on the first softirq_raise and remains unchanged until the corresponding softirq handler runs. This allows the metric to accurately reflect the maximum latency between the initial hard interrupt and the subsequent softirq execution.

go test：

functional test：
Enabled the TIMER softirq metric and extended the latency range for validation.
A custom kernel module was used to block softirq processing for 50 ms, and the collected TIMER latency met expectations.

fix: correct softirq_latency metric behavior

softirq_raise can be triggered by hard interrupts and is re-entrant.
The previous implementation updated the timestamp on every raise, causing the reported value to reflect the minimum latency, which is misleading and of limited usefulness.

Signed-off-by: lizhong [email protected]