xquery4 2.1.0

This package has no dependencies.

Minor rather than patch: four error codes change, and `fn:implicit-timezone` becomes stable
within a query. Takes **PhoenixmlDb.Core 2.0.0** (Core runs its own cadence and did not move).

Part of the coordinated 2.1.0 train with **PhoenixmlDb.Xslt 2.1.0**. The two halves of the
XPath-evaluation performance work land together — see *Performance* below for why that pairing
is the number worth quoting.

### Behaviour changes — read this before upgrading

**1. `fn:implicit-timezone` is now stable within a query.** It read `DateTimeOffset.Now` on every
call, so two calls in one query could disagree, and it could disagree with `fn:current-dateTime`
in the same query. It now reads the instant the dynamic context captured, as a dynamic-context
component is specified to. A query that was (accidentally) observing clock drift between calls
will see one value.

**2. Four error codes are now the ones the specs name.** A query that catches by code must be
updated; a query that merely fails still fails.

| condition | was | now |
|---|---|---|
| a library module declares a **function** outside its target namespace | *(no error)* | `XQST0048` |
| a module namespace with a malformed percent escape | *(no error)* | `XQST0046` |
| an ambiguous leading lone `/` | wrong code, from the wrong exception type | `XPST0003` |
| a bad escape inside `<fn:string escaped="true">` (`fn:xml-to-json`) | `FOJS0006` | `FOJS0007` |

The first two were **not raised at all** — a library module declaring a function in another
namespace was imported silently. Code relying on that import now gets a static error.

### Added

- **`QueryExecutionContext.DynamicCallDepth`** lets a host function tell that it is running
 inside a dynamic function call. This exists for XSLT: `fn:current-output-uri()` must return the
 empty sequence during a dynamic call (spec bug 30411), and the `AbsentFocus` sentinel could not
 carry that signal because the XSLT engine pushes the same sentinel for an initial template with
 no context item.

### Performance

XPath evaluation now pays per-call-site and per-context setup once instead of per evaluation:
call sites cache what they resolved (revalidated against a `FunctionLibrary.Version` counter, so
a plan shared across executions or threads cannot use another library's function), context state
is created on first use, the three focus stacks become one, and atomization stops allocating a
closure per call.

ISO Schematron three-step compile of 77 state schematrons, sequential:

| | Total | Include step |
|---|---|---|
| 2.0.0 | 5.93 min | 3.81 min |
| this release alone | 4.99 min | 2.98 min |
| **with PhoenixmlDb.Xslt 2.1.0** | **3.97 min** | **2.06 min** |

The last row is what a consumer on the 2.1.0 train actually gets: **46% off the include step.**
Either half alone gets roughly half of it, which is why these two releases are cut together.

All 231 compile-step outputs are byte-identical to 2.0.0's.

### Conformance

**W3C QT3 29,822 / 31,379 (95.0%)**, up from 29,813 on 2.0.0. The nine are the error-code fixes
above landing on cases that assert the code.

The performance work moved nothing: measured on `41cea23` with it and on `0cd012a` without it,
both runs report the same 29,822 / 31,379 and the same 1,557 failures. That is the result a
change about allocation should produce, and it was measured here rather than taken from the
branch that claimed it.

XQuery unit suite: **1,856 passed, 0 failed** on `net10.0`.