Normalize advisor observation contextual names

[AzazelSensei]

Summary

• Strip whitespace from advisor names in DefaultAdvisorObservationConvention.getContextualName() before camel-case concatenation
• Prevents stray spaces in span names such as tool _calling for Tool Calling Advisor
• Adds a focused unit test covering the Tool Calling Advisor case

Fixes #6750

Test plan

• ./mvnw -pl spring-ai-client-chat -am -Dtest=DefaultAdvisorObservationConventionTests -Dsurefire.failIfNoSpecifiedTests=false test
• Verify OTLP / Tempo span name is tool_calling with tool callbacks enabled

[kycasdzxc]

Ran the reported input through ParsingUtils.CAMEL_CASE to check the surrounding behaviour, and two things came up that might be worth folding in.

StructuredOutputValidationAdvisor hits the same bug and isn't covered.

It is the other advisor in the codebase that overrides getName() with a display string:

// StructuredOutputValidationAdvisor.java:109
return "Structured Output Validation Advisor";

On main that produces:

splitCamelCaseToLower -> ["structured ", "output ", "validation ", "advisor"]
join("_")             -> "structured _output _validation _advisor"
.replace("_advisor")  -> "structured _output _validation "

This PR fixes it correctly (structured_output_validation), so this is only a suggestion to add it to the new test. It is the second of the two real instances, and pinning both makes the regression boundary explicit.

The _around_advisor branch is still untested.

The only contextual-name test on main is contextualName() with "MyName" -> "my_name", which reaches neither replace. The test added here is the first to exercise _advisor. Since trimAllWhitespace changes what reaches both, pinning _around_advisor as well would be cheap.

One limitation worth recording, though I do not think it needs to block this: trimAllWhitespace only recovers an identifier when every word is capitalized. "Simple Logger advisor" becomes simple_loggeradvisor: no stray spaces, so the reported bug is fixed, but the suffix no longer strips. This is not a regression, since main gives simple _logger advisor for the same input, equally unstripped. It is the same assumption in a narrower form, in that getContextualName() expects Advisor#getName() to be identifier-shaped. That is the open question from #6750, and this change does not settle it either way, so it is likely a maintainer call.

[kycasdzxc]

Reworded my earlier notes as suggestions so they're easier to apply. Test-only, and entirely up to you whether to take them.

[AzazelSensei]

Thanks for the suggestions. Pushed the two extra tests (Structured Output Validation Advisor + AroundAdvisor suffix).

[kycasdzxc]

@AzazelSensei Glad they were useful. Since the suggestions were applied manually,
the co-author trailer didn't get added automatically. Could you amend
the commit to include it?

Co-authored-by: kycasdzxc <105793579+kycasdzxc@users.noreply.github.com>

[AzazelSensei]

Amended — added your Co-authored-by.