Attribute wrapped blob-hydration failures to their inner cause

[tyrielv]

What

Attribute a wrapped blob-hydration failure to its real inner cause when categorizing telemetry.

RetryableException carries the real cause in InnerException. The BlobHydrationFailureCategory
logic added in #2071 checked the RetryableException type itself:

category = errorArgs.Error is IOException ? LocalIO : NetworkUnavailable;

RetryableException is not an IOException, so every RetryableException was tagged
NetworkUnavailable - even when it wrapped a local disk/IO cause. This over-counts the
"external / network" slice and under-counts LocalIO.

This matters because RetryableException is the largest blob-hydration failure bucket in the
field (~255 os.2020 mount machines, ~2.85M events in 30 days).

Change

• Unwrap RetryableException.InnerException before categorizing.
• Map IOException / UnauthorizedAccessException / Win32Exception (the local disk/IO family) to
  LocalIO.
• A RetryableException whose inner cause is not local (e.g. HttpRequestException), or that has no
  inner cause, stays NetworkUnavailable.
• Telemetry metadata only. No behavior change. The Warning-while-retrying / Error-on-terminal
  split is unchanged.

Which path actually reaches this branch

On this branch the RetryableException arrives from Context.Repository.TryCopyBlobContentStream -
typically StreamUtil wrapping an IOException while reading a corrupt or truncated local loose
object (or the ProjFS write callback failing). Note: a RetryableException thrown while writing
a loose object (e.g. a full disk during download) is caught by the inner download retrier and
surfaces as DownloadFailed (no exception on the outer handler), so it does not reach this
branch. UnauthorizedAccessException / Win32Exception are mapped to LocalIO because they belong
to the same local disk/IO family.

Testing

• Full unit suite: 898 passed, 0 failed (11 pre-existing skips).
• Per-cause tests: IOException, UnauthorizedAccessException, Win32Exception -> LocalIO;
  HttpRequestException (non-local inner) and a bare no-inner RetryableException ->
  NetworkUnavailable.
• The fixture resets the process-global RetryCircuitBreaker in both SetUp and TearDown so
  these failure-driving tests cannot open the circuit for one another or leave it dirty for a later
  fixture.

Rollout note (metadata distribution)

This moves a large volume of failures from NetworkUnavailable to LocalIO. The dashboard folds
both into "external", so the external-vs-actionable split is unaffected, but any per-category
NetworkUnavailable / LocalIO trend or alert will show a one-time step change at the deploy
boundary. Annotate the boundary and re-baseline per-category thresholds before using this data for a
release-readiness comparison. Fully reversible (pure metadata).

Relationship to #2071

Follow-up to #2071 (now merged), which introduced BlobHydrationFailureCategory. This branch has
been rebased onto master, so the diff is just the unwrap change (two files).

Companion

Consumes the BlobHydrationFailureCategory tag from #2071. The dashboard-side split that reads these
tags is EngSys PR 16223444 in devprod.git.telemetry.