tpl/partials: Make sure a cached partial is invoked only once

This commit revises the locking strategy for `partialCached`. We have added a benchmark that may be a little artificial, but it should at least show that we're not losing any performance over this:

```bash
name              old time/op    new time/op    delta
IncludeCached-10    12.2ms ± 2%    11.3ms ± 1%   -7.36%  (p=0.029 n=4+4)

name              old alloc/op   new alloc/op   delta
IncludeCached-10    7.17MB ± 0%    5.09MB ± 0%  -29.00%  (p=0.029 n=4+4)

name              old allocs/op  new allocs/op  delta
IncludeCached-10      128k ± 1%       70k ± 0%  -45.42%  (p=0.029 n=4+4)
```

This commit also revises the template metrics hints logic a little, and add a test for it, which output is currently this:

```bash
 cumulative       average       maximum      cache  percent  cached  total
       duration      duration      duration  potential   cached   count  count  template
     ----------      --------      --------  ---------  -------  ------  -----  --------
      163.334µs     163.334µs     163.334µs          0        0       0      1  index.html
       23.749µs       5.937µs      19.916µs         25       50       2      4  partials/dynamic1.html
        9.625µs       4.812µs        6.75µs        100       50       1      2  partials/static1.html
        7.625µs       7.625µs       7.625µs        100        0       0      1  partials/static2.html
```

Some notes:

* The duration now includes the cached invocations (which should be very short)
* A cached template gets executed once before it gets cached, so the "percent cached" will never be 100.

Fixes #4086
Fixes #9506

1 files changed, 12 insertions, 0 deletions

diff --git a/tpl/math/math.go b/tpl/math/math.go
index d70a3f833..cd4a47eca 100644
--- a/tpl/math/math.go
+++ b/tpl/math/math.go
@@ -17,6 +17,7 @@ package math
 import (
 	"errors"
 	"math"
+	"sync/atomic"
 
 	_math "github.com/gohugoio/hugo/common/math"
 
@@ -162,3 +163,14 @@ func (ns *Namespace) Sqrt(a interface{}) (float64, error) {
 func (ns *Namespace) Sub(a, b interface{}) (interface{}, error) {
 	return _math.DoArithmetic(a, b, '-')
 }
+
+var counter uint64
+
+// Counter increments and returns a global counter.
+// This was originally added to be used in tests where now.UnixNano did not
+// have the needed precision (especially on Windows).
+// Note that given the parallel nature of Hugo, you cannot use this to get sequences of numbers,
+// and the counter will reset on new builds.
+func (ns *Namespace) Counter() uint64 {
+	return atomic.AddUint64(&counter, uint64(1))
+}