Adds optional VCPU (int) and MemoryGB (float64) fields to
common.ComputeDetails so per-provider catalogue lookups can enrich
compute recommendations with sizing data. Both fields use json
omitempty — converters that don't yet wire a catalogue leave them at
the zero value and the API payload stays clean.

GetDetailDescription appends " (N vCPU / M GB)" when both values are
> 0, otherwise returns the existing "platform/tenancy" form. Format
uses %g so 16 GB renders as "16 GB" (not "16.000000 GB") while 0.5 GB
SKUs render as "0.5 GB".

Scope decision — schema-only foundation:

- Azure: PR #81 (perf/azure-batched-sku-lookup) introduces the
  cachedSKULookup helper for cache/cosmos/database. Compute was
  scoped back from #81 because these fields didn't exist. With the
  schema in place, a follow-up PR can wire compute on top of #81 once
  it merges (filed as a separate issue post-merge to avoid an
  unmerged-PR dependency in this changeset).
- AWS: ce.EC2InstanceDetails (Cost-Explorer) has no vCPU/Memory
  fields; populating would require a new ec2:DescribeInstanceTypes
  caller + cache + mocks. Substantial net-new code, tracked as a
  separate follow-up.
- GCP: no compute recommendation converter exists today. N/A.
- Frontend: api.Recommendation has no `details` field and the detail
  drawer renders a hard-coded list. Surfacing the new fields needs
  separate API + UI work; tracked as follow-up once at least one
  provider populates the fields.

Tests: extend pkg/common/types_test.go::TestComputeDetails_GetDetail
Description with 4 new table rows — VCPU-only, MemoryGB-only,
integer-GB, and fractional-GB (Azure 0.5 GB SKU shape).

Refs #82

closes #148) (#229)

PR #97 added VCPU + MemoryGB fields to common.ComputeDetails (with
omitempty JSON tags). PR #81 introduced the lazy SKU-catalogue pattern
for cache/cosmosdb/database but explicitly scoped back compute because
the destination fields didn't exist. Both prerequisites are now in
place; this wires Azure compute to the same pattern.

Implementation mirrors cache/database verbatim for consistency:

- New unexported vmSKUEntry{vCPUs, memoryGB}.
- New skuCacheOnce sync.Once + skuCacheMap field on ComputeClient.
- cachedSKULookup(ctx, skuName) lazily triggers the catalogue fetch on
  first call; subsequent calls are O(1) map lookups. ok=false on miss
  or fetch error.
- fetchSKUCatalogue reuses the existing createResourceSKUsPager and
  isAvailableInRegion helpers (already used by GetValidResourceTypes),
  walks every page, and reduces virtualMachines SKUs into the map.
  Returns nil on error so the sync.Once-gated cache stays nil and
  converters fall back to the empty-fields path with a one-time WARN.
- extractVMSKUCapabilities pulls vCPUs (Atoi) and MemoryGB
  (ParseFloat) from the SKU's Capabilities name/value list. Unparseable
  or missing capabilities → 0, treated as "unknown".
- populateVMSKUMapFromPage was extracted out of fetchSKUCatalogue to
  stay under the project's gocyclo=10 threshold (matches the
  cache/database extraction pattern from PR #81).
- convertAzureVMRecommendation now takes ctx (was _) and populates
  Details.VCPU/MemoryGB from the cache when both >0. The single caller
  GetRecommendations already passes ctx; no public API change.

Behaviour preserved on failure: a transient ResourceSKUsClient error
no longer breaks the conversion — VCPU/MemoryGB stay at 0 (omitempty
hides them from API payloads), the rest of Details is populated from
the recommendation payload as before, and a WARN is logged once per
client lifetime.

UX improvement: common.ComputeDetails.GetDetailDescription appends
" (<vcpu> vCPU / <memory> GB)" when both fields are >0, so Azure VM
recommendation summaries now include the size string instead of the
SKU name alone.

Tests added in providers/azure/services/compute/client_test.go
(file-scoped vmSKUCatalogueMockPager keeps the shared
mocks.MockResourceSKUsPager surface untouched, mirroring the
cosmosdb / cache test convention):

- TestComputeClient_ConvertAzureVMRecommendation_PopulatesVCPUAndMemoryFromSKUCache:
  catalogue hit populates VCPU=2, MemoryGB=8 for Standard_D2s_v3.
- TestComputeClient_ConvertAzureVMRecommendation_PagerErrorFallsBack:
  fetch error leaves both fields at 0; conversion succeeds.
- TestComputeClient_ConvertAzureVMRecommendation_NoMatchLeavesFieldsZero:
  catalogue miss leaves both fields at 0; conversion succeeds.
- TestComputeClient_CachedSKULookup_FetchedOnce: 10 lookups trigger
  exactly 1 NextPage call (pins the N+1 invariant per PR #81).

Out of scope (separate follow-ups, per the issue body):

- AWS / GCP compute converter wiring.
- Platform / Tenancy / Scope enrichment (require non-ResourceSKUs
  Azure data sources).

Verification: gofmt clean; go vet ./... clean; gocyclo -over 10 finds
no offenders in the modified file; go test
github.com/LeanerCloud/CUDly/providers/azure/services/compute → 42
passing (4 new + all existing). The pre-existing
providers/azure/services/search build failure (missing go.sum entry)
is unrelated to this change.

Closes #148